Request: ROCm Support for fairseq2 (PyTorch ROCm builds)

**Is your feature request related to a problem? Please describe:**
Currently, apart from fairseq, fairseq2 provides pre-built wheels only for CPU and CUDA backends. There is no option for users running PyTorch with ROCm, which means fairseq2 cannot be installed or used on AMD GPUs. This restricts the usability of fairseq2 for researchers and developers using ROCm environments with AMD GPUs.

**Describe the solution you would like:**
Add ROCm-compatible builds of fairseq2.

The issue names no files, tests, or entry points. Start by locating the packaging and build or CI configuration for the existing CPU and CUDA wheels, then compare how PyTorch backends are selected. Done means fairseq2 has installable ROCm-compatible builds for AMD GPU environments, with validation alongside the existing wheels.

Written by the indexing model from the issue text.