[release] Gate the CUDA wheel on vllm._C_stable_libtorch

Upstream moved the CUDA custom ops out of the monolithic vllm._C into
vllm._C_stable_libtorch (and _moe_C_stable_libtorch), registered through
STABLE_TORCH_LIBRARY; CMakeLists now builds the legacy _C target for ROCm
only. A correct CUDA wheel therefore ships _C_stable_libtorch.abi3.so and no
_C.abi3.so, which the release gate rejected as "required wheel contents are
absent: {'vllm._C': 'absent'}" and the on-GPU probe would fail importing
vllm._C.

Point REQUIRED_EXTENSIONS, the packaged-contents check, the runtime gate, and
the validation import probe at vllm._C_stable_libtorch. Refresh the release-gate
test fixtures to the cu130 / torch 2.13.0 toolchain the build now uses so they
match config.json.

[release] Publish paired TPU wheel candidates

[release] Enforce exact TPU source checkout

[release] Skip GPU builds for TPU-only changes

[release] Publish paired TPU wheel candidates

[release] Isolate wheel validation from source checkouts (#32)

Co-authored-by: rjpower <rjpower@users.noreply.github.com>

[release] Isolate wheel validation from source checkouts (#32)

Co-authored-by: rjpower <rjpower@users.noreply.github.com>

[release] Export only the native wheel from BuildKit (#31)

Co-authored-by: rjpower <rjpower@users.noreply.github.com>

[grugmoe] Fix mixed-attention cache configuration (#13)

Co-authored-by: yoblin <268258002+yoblin@users.noreply.github.com>

Merge upstream/main into fork main (sync 2026-05-23)