Fine-tunes the PI0.5 Vision-Language-Action model on a LeRobot robotics
dataset stored in S3, using Ray Data for CPU preprocessing and Ray Train
for distributed GPU training.

Key features:
- Streams LeRobot v3 (parquet + mp4) data from S3 via a custom Ray Data
  datasource with anonymous S3 access
- Preprocesses on CPU workers (rename cameras, HWC->CHW, /255 normalise)
  so GPU workers are never blocked on I/O or video decoding
- Expert-only fine-tuning: only the 4 action/time projection heads are
  trained; the PaliGemma backbone stays frozen
- BF16 mixed precision throughout (no GradScaler needed)
- Linear-warmup + cosine-decay LR schedule
- Fault-tolerant checkpointing via ray.train.report()
- Declarative, cloud-agnostic compute config targeting 8x L40S GPUs

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Signed-off-by: Robert Nishihara <rkn@anyscale.com>