ci: expand ROCm architecture coverage

[sstamenk]

Summary

• add RDNA1 targets gfx1010, gfx1011, and gfx1012 to the ROCm 7.14 Linux and Windows artifacts
• add gfx908 and gfx1030 to every supported Linux ROCm build
• align the Windows ROCm 7.14 artifact with AMD's multi-architecture index by adding gfx908, gfx90a, and gfx1103
• update the README platform summary and installation target matrix to document the expanded coverage

Architecture additions

ROCm	Linux additions	Windows additions
6.4.4	gfx908, gfx1030	Not built
7.0.2	gfx908, gfx1030	Not built
7.1.1	gfx908, gfx1030	Not built
7.2.1	Not built	No change; existing targets match the PyTorch wheel
7.2.4	gfx908, gfx1030	Not built
7.14.0	gfx1010, gfx1011, gfx1012	gfx908, gfx90a, gfx1010, gfx1011, gfx1012, gfx1103

Validation

• pre-commit run --all-files
• bash -n .github/scripts/build-rocm.sh
• compared ROCm 7.14 targets with the platform-specific device wheels in https://repo.amd.com/rocm/whl-multi-arch/
• downloaded and inspected all 207 HIP fat-binary bundles in the Windows ROCm 7.2.1 PyTorch wheel; confirmed its existing bitsandbytes target list already exactly matches gfx1100, gfx1101, gfx1102, gfx1150, gfx1151, gfx1200, and gfx1201

[sstamenk]

One thing I'm not sure is whether to include the CDNA1 gfx908
gfx90a architectures in the ROCm 7.14 Windows build. The https://repo.amd.com/rocm/whl-multi-arch/ index has the wheels for these but not other CDNA architectures. Also the Windows builds only included RDNA until this point.

[github-actions]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.