Implement PyTorch RLDS data loader with distributed support

[tahsinkose]

Summary

Implements the PyTorch RLDS data loader, replacing the NotImplementedError stub in
create_rlds_data_loader. This enables PyTorch DDP training on RLDS-format datasets (e.g.,
DROID).

Design

• Rank-0-only pipeline: Only rank 0 builds the heavy TF-based RLDS pipeline; other ranks
  receive batches via torch.distributed.broadcast to avoid redundant resource usage.
• Broadcast + shard: Each full batch is broadcast from rank 0, then sliced so every rank gets
  batch_size // world_size samples.
• Exhaustion sync: A broadcast flag ensures all ranks break together, preventing deadlocks on
  collectives.
• Single-GPU: When torch.distributed is not initialized, the loader degenerates to a simple
  pass-through with no overhead.