Train fallback model

## Summary
Train the neural fallback model on the training dataset.

## Success Criteria
- [ ] Training loop implemented with proper logging
- [ ] Validation loss tracked
- [ ] Best checkpoint saved automatically
- [ ] Training reproducible with fixed seeds
- [ ] Training completes in reasonable time (<24h on consumer GPU)
- [ ] Final model meets latency requirements

## Training Details
- Loss: Cross-entropy on next token prediction
- Optimizer: AdamW
- LR schedule: Warmup + cosine decay
- Batch size: Tune for hardware

## Experiment Tracking
- Log to experiments/neural-fallback-v1/
- Track: loss curves, learning rate, gradient norms
- Save: config, final metrics, best checkpoint

## Hardware Target
- Should train on single consumer GPU (RTX 3080 or similar)
- CPU training acceptable if slow but completes