Add optional early stopping to training loop

[AkilsuryaS]

Summary

This PR adds optional early stopping support to train.py while keeping the default training behavior unchanged.

What changed

• Added configurable early-stopping knobs:
  • ENABLE_EARLY_STOP
  • EVAL_EVERY
  • EARLY_STOP_PATIENCE
  • MIN_DELTA
  • TARGET_VAL_BPB
  • SAVE_BEST_CHECKPOINT
  • BEST_CKPT_PATH
• Added periodic validation during training when early stopping is enabled.
• Added stop conditions for:
  • target validation BPB reached
  • patience exhausted (no sufficient improvement)
• Added optional best-checkpoint save/restore path for final evaluation.
• Added richer end-of-run reporting for early-stop metrics (best_val_bpb, best_step, last_val_bpb, stopped_early, stop_reason) and startup_seconds.

Behavior

• Default flow is unchanged (ENABLE_EARLY_STOP = False).
• Existing time-budget stop remains in place.
• New logic activates only when early stopping is explicitly enabled.