fbgemm-xpu: add CI workflow

[flezaalv]

Adds .github/workflows/ci-fbgemm-xpu.yml, a dedicated CI workflow for fbgemm-xpu so it can be built, tested and released independently of torchcodec-xpu.

• Builds manylinux wheels for Python 3.10–3.14 using pytorch/manylinux2_28-builder:xpu-v2.12.0-rc1 (ships oneAPI + Intel driver stack).
• Composes per-Python wheels into a single artifact and publishes to PyPI on release: published.
• Runs GPU tests on bmg and dg2 self-hosted runners with a minimal Intel runtime (libze1, libze-intel-gpu1, intel-ocloc).

Test plan

• Trigger the workflow via PR to main and verify all jobs succeed.
• Confirm build produces wheels for every Python version in the matrix.
• Confirm test-fbgemm reports torch.xpu.is_available() == True on both bmg and dg2.
• Confirm compose merges wheels correctly.
• Verify publish is gated behind the release event and skipped on PR runs.

This PR is intended as the base for #71.

cc: @aagalleg, @manuelhsantana, @luis-real