Add optional checkpoint import/export hooks

[corysmart]

Summary

This adds optional checkpoint load/save hooks to train.py without changing default behavior.

New behavior is only enabled when these env vars are set:

• AUTORESEARCH_LOAD_CHECKPOINT
• AUTORESEARCH_SAVE_CHECKPOINT

What changed

• add maybe_load_checkpoint() to load a checkpoint before training
• allow partial state-dict restore when tensor shapes still match
• add maybe_save_checkpoint() to persist model state and a small amount of run metadata after evaluation
• keep the compiled training path unchanged by loading/saving around the existing model_core setup

Why

This creates a minimal continuation/export seam that is useful for:

• resuming or bootstrapping experiments
• external harnesses that want to persist model state between short runs
• downstream tooling that needs checkpoint artifacts without restructuring the trainer

Notes

• default behavior is unchanged when the env vars are unset
• checkpoint loading is permissive by shape, so compatible subsets can still be restored
• no new dependencies were added