Classification trainer: restore optimizer/scheduler state on resume

[jeetv]

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 them into every checkpoint all along -- only the epoch counter was ever restored.

Port localization_trainer.py's build_trainer(..., resume_from=...) pattern: Trainer_Classification.load() gains resume_training=True, which builds a fresh optimizer/scheduler from config and restores their state via the existing load_checkpoint() plumbing, stashing the result in self._resume_state. Trainer_Classification.train() gains a resume_from parameter that reuses that state instead of rebuilding it, seeds best-metric tracking, continues from the saved epoch, and guards against resuming a checkpoint that already reached TRAIN.epochs. apis/classification.py wires weights= on ClassificationModel.train() to request resume_training the same way LocalizationModel.train() already does for weights= on its own train().

No behavior change for fresh training (resume_from=None keeps the original build_optimizer()/build_scheduler() path byte-for-byte) or for load_weights() /infer() (resume_training defaults to False there too).

Adds tests/test_classification_trainer_resume.py, which trains a real epoch, saves, and asserts the resumed Adam/StepLR state_dict matches the checkpoint rather than a freshly-initialized optimizer.