fix: defer np.typing annotation evaluation

[xyf5432]

Fixes #127

Problem

np.typing.XXX is only accessible on numpy < 2.0 if numpy.typing has been explicitly imported earlier. Without a version constraint forcing numpy >= 2, users with numpy 1.x installed can hit:

AttributeError: module 'numpy' has no attribute 'typing'

when the annotation is evaluated at import time.

Fix

File: ffn/training/examples.py

+ from __future__ import annotations
+
  def update_seeds(self, batched_seeds: np.typing.ArrayLike):

(Optional follow-up: the annotation is also narrower than the documented contract — the docstring says the argument may be an "array-like object backed by accelerator memory" (i.e. a JAX array), which np.typing.ArrayLike does not formally cover. np.typing.ArrayLike | jax.typing.ArrayLike would describe the accepted inputs more accurately.)

Verification

• Verified the failure mechanism in isolation on numpy 1.26.4: np.typing.ArrayLike raises AttributeError: module 'numpy' has no attribute 'typing' unless numpy.typing has been imported first. Contrast tests confirmed that importing pandas 3.x or xarray beforehand resolves it, while importing scipy does not — i.e. whether training crashes today depends on import order, which this fix eliminates.
• With from __future__ import annotations, the annotation is no longer evaluated when BatchExampleIter is defined, making the training entry points (train.py, JAX path) safe on every numpy version.
• The change is annotation-only — no behavioral change.

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.