Update wheels matrix for ROCM + README update

wheels/windows timeout (#1309)

* wheels/windows timeout

Try building with `MAX_JOBS=3`

* Update wheels_build.yml

Bump version to 0.0.32 (fairinternal/xformers#1418)

__original_commit__ = fairinternal/xformers@dee2011

remove merge_attentions backward (fairinternal/xformers#1402)

__original_commit__ = fairinternal/xformers@601197a

Build for PyTorch 2.7.1

ghstack-source-id: 687672fe86b9b7db141537adc27658fa9bb245e7
Pull-Request: fairinternal/xformers#1399

__original_commit__ = fairinternal/xformers@96f264c

Version 0.0.30

ghstack-source-id: b4660792cfd868da3d31a685bc7a43d3679b77c9
Pull Request resolved: fairinternal/xformers#1354

__original_commit__ = fairinternal/xformers@80dc26c

Fix missing release for CUDA 12.6 on Windows (#1217)

Fixes #1215

Build with PyTorch 2.6.0, update CUDA targets (#1204)

Fix number of return values for Flash2 (fairinternal/xformers#1279)

Probably we should have a test somewhere, but it becomes hard to cover all combinations of {flash version}x{GPU}x{PyTorch version}x{Built-in Flash / PT flash / third-party flash}

__original_commit__ = fairinternal/xformers@9c39da4

Changelog for 0.0.29 (fairinternal/xformers#1275)

__original_commit__ = fairinternal/xformers@b6c5e6c