Upgrade colab base image to 20260917-060051 (#1562)

### Base image
- GPU and CPU images now use
`release-colab-external-images_20260917-060051_RC00` (previously
`20260716-060051`).
- The new base image moves from Python 3.12 to Python 3.13.

### Dockerfile.tmpl
- **Python paths:** `PACKAGE_PATH` and the `sitecustomize.py`
destination now point to `python3.13` instead of `python3.12`.
- **Stopped installing the `keras_internal.py` patch:**
`tensorflow_decision_forests` is no longer in the base image, and its
latest release (1.12.0) has no cp313 wheel, so the
`tensorflow_decision_forests/keras/` directory it patched doesn't exist.
- **vtk apt dependency:** replaced `libgl1-mesa-glx`, which newer Ubuntu
no longer provides, with `libgl1` and `libglx-mesa0`.

### kaggle_requirements.txt
**Removed**
- `tensorflow-io`: the last release (0.37.1, July 2024) requires
`python<3.13`, has no cp313 wheels, and the project is unmaintained.
- The `jupyter-lsp==1.5.1` pin (b/276358430): on 3.13 it keeps
jupyterlab on 3.x, which pulls in `jupyter-ydoc 0.2.5` and then `y-py
0.6.2`. `y-py` has no cp313 wheel and its sdist no longer builds.
jupyterlab 4.x uses `pycrdt` instead, so jupyter-lsp is now unpinned.
- The `fastcore<2` pin and the `python-fasthtml<0.14.6` pin that only
existed because of it (b/545164527).

**Added or changed**
- `opencv-python<5`, `opencv-python-headless<5` and
`opencv-contrib-python<5`: the base image mixes OpenCV versions
(`opencv-python`/`-headless` at 5.0.0.93, `opencv-contrib-python` at
4.14.0.94). All three install into the same `cv2/` package and overwrite
each other, and OpenCV 5.0 drops long-standing APIs such as
`cv2.CascadeClassifier`. Holding all three on 4.x keeps `cv2` consistent
and backward compatible.
- `flax` is now `flax>=0.12`.
- Added `universal-pathlib` because Pyxis
([#1402](Kaggle/kaggle-environments#1402),
merged Sep 18) added `universal-pathlib` to kaggle-environments'
[pyproject.toml](https://github.com/Kaggle/kaggle-environments/blob/d7729da06cc1382eb742d6980dc3180aa85caa28/pyproject.toml#L41).

b/493600019

Author: tifftoff

Dockerfile.tmpl

@@ -1,7 +1,7 @@
 {{ if eq .Accelerator "gpu" }}
-FROM us-docker.pkg.dev/colab-images/public/runtime:release-colab-external-images_20260716-060051_RC00
+FROM us-docker.pkg.dev/colab-images/public/runtime:release-colab-external-images_20260917-060051_RC00
 {{ else }}
-FROM us-docker.pkg.dev/colab-images/public/cpu-runtime:release-colab-external-images_20260716-060051_RC00
+FROM us-docker.pkg.dev/colab-images/public/cpu-runtime:release-colab-external-images_20260917-060051_RC00
 {{ end}}
 
 ADD kaggle_requirements.txt /kaggle_requirements.txt
@@ -36,7 +36,7 @@ ADD clean-layer.sh  /tmp/clean-layer.sh
 ADD patches/nbconvert-extensions.tpl /opt/kaggle/nbconvert-extensions.tpl
 ADD patches/template_conf.json /opt/kaggle/conf.json
 
-ARG PACKAGE_PATH=/usr/local/lib/python3.12/dist-packages
+ARG PACKAGE_PATH=/usr/local/lib/python3.13/dist-packages
 
 # Install GPU-specific non-pip packages.
 {{ if eq .Accelerator "gpu" }}
@@ -56,9 +56,9 @@ RUN sed -i "s/httpredir.debian.org/debian.uchicago.edu/" /etc/apt/sources.list &
     apt-get install -y graphviz && pip install graphviz && \
     /tmp/clean-layer.sh
 
-ADD patches/keras_internal.py \
-    patches/keras_internal_test.py \
-    $PACKAGE_PATH/tensorflow_decision_forests/keras/
+# patches/keras_internal.py is not installed right now. tensorflow_decision_forests
+# is no longer in the base image and its latest release (1.12.0) ships no cp313
+# wheel, so there is no tensorflow_decision_forests/keras/ to patch.
 
 RUN apt-get install -y libfreetype6-dev && \
     apt-get install -y libglib2.0-0 libxext6 libsm6 libxrender1 libfontconfig1 --fix-missing && \
@@ -86,7 +86,7 @@ RUN mkdir -p /usr/share/nltk_data && \
 # git-lfs 3.7.1 built with Go 1.26, and Ubuntu's git-lfs 3.0.2 package
 # overwrites it with an older build from Go 1.18.
 # vtk dependencies
-RUN apt-get install -y libgl1-mesa-glx && \
+RUN apt-get install -y libgl1 libglx-mesa0 && \
     # xvfbwrapper dependencies
     apt-get install -y xvfb && \
     /tmp/clean-layer.sh
@@ -147,7 +147,7 @@ ADD patches/kaggle_gcp.py \
 
 # Figure out why this is in a different place?
 # Found by doing a export PYTHONVERBOSE=1 and then running python and checking for where it looked for it.
-ADD patches/sitecustomize.py /usr/lib/python3.12/sitecustomize.py
+ADD patches/sitecustomize.py /usr/lib/python3.13/sitecustomize.py
 
 ARG GIT_COMMIT=unknown \
     BUILD_DATE=unknown

kaggle_requirements.txt

@@ -26,17 +26,10 @@ dipy
 docker
 easyocr
 emoji
-# TODO: b/545164527 - unpin once fastai supports fastcore v2.
-# fastcore v2 removed L.starmap, breaking fastai's Optimizer.set_hypers:
-# https://github.com/fastai/fastai/issues/4154
-# Pinned here rather than on a direct dep because nbdev, execnb, ghapi, fastgit
-# and fastspec all pull fastcore uncapped; only constraining fastcore itself
-# makes the resolver backtrack all of them.
-fastcore<2
 fasttext
 featuretools
 fiona
-flax
+flax>=0.12
 # TODO: b/545164527 - unpin once dipy ships its fury<2.0 cap.
 # fury 2.0 swapped VTK for pygfx/wgpu, breaking `from dipy.viz import window`.
 # dipy requires fury uncapped; fix is merged upstream but unreleased:
@@ -62,8 +55,11 @@ isoweek
 jedi
 # jitler 0.11.1 breaks simulation image
 jiter==0.10.0
-# b/276358430: fix Jupyter lsp freezing up the jupyter server
-jupyter-lsp==1.5.1
+# b/276358430: jupyter-lsp was pinned to 1.5.1 to stop Jupyter lsp freezing up
+# the jupyter server. That pin is unusable on Python 3.13: it holds jupyterlab
+# back to the 3.x line, which pulls jupyter-ydoc 0.2.5 -> y-py 0.6.2, and y-py
+# publishes no cp313 wheel and its sdist no longer builds. jupyterlab 4.x
+# replaced y-py with pycrdt, so leave jupyter-lsp unpinned.
 # b/333854354: pin jupyter-server to version 2.12.5; later versions break LSP (b/333854354)
 jupyter_server==2.12.5
 jupyter_server_proxy
@@ -91,6 +87,15 @@ nbdev
 nilearn
 olefile
 onnx
+# The base image ships an inconsistent OpenCV set: opencv-python and
+# opencv-python-headless at 5.0.0.93 but opencv-contrib-python at 4.14.0.94.
+# They all install into the same cv2/ package and clobber each other, and
+# OpenCV 5.0 dropped long-standing APIs (e.g. cv2.CascadeClassifier). Hold the
+# whole set on the last 4.x release so cv2 stays consistent and notebooks that
+# use the 4.x API keep working.
+opencv-contrib-python<5
+opencv-python<5
+opencv-python-headless<5
 openslide-bin
 openslide-python
 optuna
@@ -113,10 +118,6 @@ pyexcel-ods
 pymongo
 pypdf
 pytesseract
-# fastprogress pulls in python-fasthtml uncapped, and 0.14.6 raised its
-# floor to fastcore v2, which conflicts with the fastcore<2 pin above.
-# Remove once fastcore can be unpinned.
-python-fasthtml<0.14.6
 python-lsp-server
 pytorch-ignite
 pytorch-lightning
@@ -132,7 +133,9 @@ scikit-plot
 scikit-surprise
 git+https://github.com/facebookresearch/segment-anything.git
 squarify
-tensorflow-io
+# tensorflow-io was dropped when the base image moved to Python 3.13. Its last
+# release (0.37.1, 2024-07-01) declares requires_python <3.13 and ships no cp313
+# wheels, and the project is unmaintained.
 # Must be compatible with torch version: https://github.com/meta-pytorch/torchcodec?tab=readme-ov-file#installing-torchcodec
 torchcodec==0.11.0
 torchinfo
@@ -142,6 +145,7 @@ transformers>=5.0.0
 # The base image ships urllib3 2.5; pin up to a newer release.
 # Remove once the base image ships >=2.7 itself.
 urllib3>=2.7
+universal-pathlib
 vtk
 wavio
 xvfbwrapper