Classification trainer: restore optimizer/scheduler state on resume

[mohsayed27]

Summary

Trainer_Classification.train() always called build_optimizer()/build_scheduler() fresh, so Adam's momentum and the scheduler's LR-schedule state silently reset on every resume — even though _save_checkpoint() had been writing both into every checkpoint all along. Only the epoch counter was ever actually restored.

localization_trainer.py's build_trainer(..., resume_from=...) already solves this correctly for the localization task. This PR ports that pattern to the classification trainer:

• Trainer_Classification.load() gains resume_training=False; when True, it rebuilds the optimizer/scheduler from config and restores their state via the existing load_checkpoint() plumbing (which already supported this — it just had no caller wiring it up for classification), stashing the result in self._resume_state.
• Trainer_Classification.train() gains resume_from=None; when given, it reuses that optimizer/scheduler instead of rebuilding them, continues from the saved epoch, seeds best-metric tracking, and raises a clear error if the checkpoint already reached TRAIN.epochs (mirroring localization's guard).
• apis/classification.py wires weights= on ClassificationModel.train() to request resume_training, the same way LocalizationModel.train() already treats weights= as a resume trigger — no new public API surface.

Fresh training (resume_from=None) is unchanged — verified byte-for-byte identical code path — and so is load_weights()/inference (resume_training defaults to False there too).

Test plan

• Added tests/test_classification_trainer_resume.py: trains one real epoch with a real Adam optimizer, saves via _save_checkpoint, resumes, and asserts the resumed optimizer's state_dict()["state"] matches the checkpoint's saved momentum buffers field-by-field (while a freshly-built optimizer over the same model has empty state) — proves actual restoration, not just "doesn't crash."
• Verified the test is meaningful: reverting the source changes makes 3 of the 4 new tests fail with TypeError on the new keyword arguments, confirming they're genuinely exercising the new code.
• All 4 new tests pass; ran on the current main tip after rebasing.

🤖 Generated with Claude Code