AMD Official Support: Add ROCm Docker support for PyTorch training on AMD GPUs

[Treemann]

Summary

• Add ROCm Docker image and compose file for PyTorch training on AMD GPUs
• Document setup, workflow, and design trade-offs in docs/rocm.md
• Link from README and docs/docker.md; no changes to existing NVIDIA paths

Motivation

openpi's pyproject.toml pins CUDA builds of torch and jax, making uv sync unusable on ROCm hosts. This provides a Docker-based alternative that preserves the ROCm PyTorch stack from the base image.

Scope

• PyTorch training on AMD GPUs (tested on 8× MI308X)
• Docker-only; JAX GPU training and policy serving are out of scope

Test plan

• Image builds and passes smoke test
• 8-GPU visibility confirmed in container
• JAX → PyTorch checkpoint conversion
• Multi-GPU torchrun training on pi05_libero
• NVIDIA Docker workflow unaffected