Decouple publication plotting from Snakemake

[clairemerker]

The publication target of snakemake now only writes a manifest, figures are produced by standalone Jupyter notebooks that read that manifest.

Changes

• evalml publication <config> → now runs one cheap rule (publication_manifest) to create the manifests. Deleted the three figure rules and the python -m evalml.publication renderer CLI to simplify the code.
• Result NC files (verification, scoremaps) come from evalml experiment as before, the publication target no longer pulls them.
• New notebooks notebooks/publication/{leadtime,meteogram,scoremaps}.ipynb load the manifest via evalml.publication.resolver.Manifest, style with the packaged evalml.publication.style, and write to output/figures/<truth>/<figure>/.
• Shared style promoted into the package; plot_meteogram_region.py repointed, old marimo scripts removed.
• Added an opt-in notebooks dependency group (jupyterlab/nbconvert/ipykernel), core deps and inference/verification/hashing untouched.

Usage

evalml experiment  config/varda-single_paper_stations.yaml   # produce results
evalml publication config/varda-single_paper_stations.yaml   # write manifest

# interactive:
EVALML_MANIFEST=output/publication/<truth>/manifest.json jupyter lab notebooks/publication/

# headless (render without opening the notebook):
EVALML_MANIFEST=output/publication/<truth>/manifest.json \
  jupyter nbconvert --to notebook --execute --inplace notebooks/publication/leadtime.ipynb

See docs/publication_figures.md for the full workflow, config options, and troubleshooting.

[jonasbhend]

I like the proposed change for its simplicity. Personally, I would avoid converting the marimo notebooks to jupyter, as I fear this will be harder to work with collaboratively due to its json format and resulting git diff gibberish. Or is this a solved issue (and I am outing myself as being old :-) ).