[Bugfix] Exclude SM107 (Rubin) from support_deep_gemm()

[meena-at-work]

Summary

support_deep_gemm() uses is_device_capability_family(100), which matches
any 10.x compute capability. That silently admits SM107 into the
DeepGEMM/CuTe-gated code paths alongside SM100/SM103 -- but the vendored
DeepGEMM/CuTe kernels aren't built for native sm_107, so any call behind
is_deep_gemm_supported() (e.g. the MHC TileLang prenorm-GEMM path) hits a
cuModuleLoadData -> CUDA_ERROR_ASSERT at kernel load time on that hardware.

This replaces the family-wide check with explicit is_device_capability(100)
/ is_device_capability(103) entries, matching how Hopper (90) and the
120 family are already enumerated individually in this same function. When
DeepGEMM ships kernels built for native sm_107, it can be added back the
same way -- this isn't a permanent hardware exclusion, just a "not built
yet" gap.

Why this isn't a duplicate

Searched open PRs referencing support_deep_gemm/DeepGEMM SM gating. The
closest is #53055 ("Guard DeepGEMM in mhc_pre_broadcast_tilelang with a
torch fallback"), which is a different bug: it adds a missing
is_deep_gemm_supported() check at one call site for platforms where
DeepGEMM isn't installed at all (SM121/GB10). This PR instead fixes the
shared platform-level gate itself, which currently wrongly reports support
on hardware where DeepGEMM is installed but not built for the native arch.

Test plan

Validated live on real SM107 hardware, using the public
vllm/vllm-openai:cu134-nightly image (no NVIDIA-internal patches):

>>> from vllm.platforms import current_platform
>>> current_platform.get_device_capability()
DeviceCapability(major=10, minor=7)
>>> current_platform.support_deep_gemm()
True      # before this change

After applying this diff to the installed vllm/platforms/cuda.py in the
same container:

>>> current_platform.support_deep_gemm()
False     # after this change

Confirms the gate now correctly excludes SM107 while leaving SM100/SM103/
Hopper/120-family behavior unchanged (traced statically; no regression
hardware available to re-verify those paths, but the change is additive/
narrowing only for SM107 and doesn't touch their branches).

AI assistance disclosure: this change was drafted with Claude Code
assistance; I reviewed the diff and reasoning end-to-end, and validated the
behavior change live on SM107 hardware myself before submitting.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[meena-at-work]

@tlrmchlsmth -- please review this minor SM107 related change, this affects GLM-5.3-flash.

[meena-at-work]

@wangshangsam @xinli-sw please review and merge.

[meena-at-work]

Closing this bugfix in favour of: #59503, since the current failure mode is only on the JIT path.