Use local rank for distributed Torch CUDA binding

[sylvesterkaczmarek]

Summary

Bind each distributed Torch worker to its node-local CUDA device instead of using the worker's global distributed rank as a device index.

init_distributed() currently uses RANK both for process-group identity and torch.cuda.set_device(). That works on a single node, but in a standard multi-node torchrun launch global ranks continue across nodes while CUDA device indices restart at zero on each host.

Fixes #266.

Fix

• keep global RANK for dist.init_process_group();
• read LOCAL_RANK for CUDA device selection;
• fall back to global RANK when LOCAL_RANK is absent, preserving existing single-node/custom-launch behavior;
• keep output suppression keyed to global rank so only global rank zero prints normally.

Regression coverage

Adds a CPU-only mocked regression with RANK=5, LOCAL_RANK=1, and WORLD_SIZE=8 verifying that:

• process-group initialization receives global rank 5;
• CUDA binding and warmup use cuda:1;
• output suppression still receives global rank 5.

No tensor-parallel math, NCCL warmup behavior, or process-group configuration is otherwise changed.