[moe training] Support cutlass-dsl 4.6 in the CuTe DSL utils

[wolfcomos]

Stack (emulated ghstack, oldest at bottom):

• [nvfp4 training][rl] Add the four-over-six grouped GEMM and dispatcher wiring #14
• [nvfp4 training][rl] Add a CuTe DSL fast-path kernel for four-over-six quantize #13
• [nvfp4 training][rl] Add the four-over-six NVFP4 dense training recipe #12
• -> [moe training] Support cutlass-dsl 4.6 in the CuTe DSL utils #11

Summary

cutlass-dsl 4.6.0 dropped the leading result-type argument from the nvvm wrapper calls (cvt_packfloat, cvt_packfloat_f32, fmax) that the MXFP8 CuTe DSL utils emit, so kernels built against the <= 4.5.x call form fail to trace on 4.6 images. This PR branches on the installed cutlass-dsl version and emits the matching call form; the version probe (importlib.metadata.version("nvidia-cutlass-dsl")) treats PackageNotFoundError as the current API rather than failing the whole package import.

Design notes

• Container image churn ships cutlass under changed API and distribution names: the runtime availability probe checks module importability, not distribution names, so cutlass can be importable while the nvidia-cutlass-dsl distribution name is absent (re-vendored under another name). The version probe must not crash unrelated moe_training imports in that case; it assumes the current (4.6+) API and lets tracing surface any residual mismatch.
• The version branch is evaluated once at module scope, next to the existing runtime-availability gate, so per-op call sites stay unchanged.

Test plan

CPU (no GPU; import + collection inside the CI-equivalent container, this branch first on PYTHONPATH):

python3 -c "import torchao.prototype.moe_training.kernels.mxfp8.cute_utils"  # OK
python3 -m pytest test/prototype/moe_training/ --collect-only -q
# 427 tests collected

Lint: ruff check and ruff format --check (ruff 0.11.6, the version pinned by .github/workflows/ruff_linter.yml) pass on the changed file.

GPU results come from the assembled stack tip on GB200 (full moe_training GPU suite: 207 passed / 0 failed), since ao CI has no SM100 runner. Runtime GPU evidence comes from cutlass-dsl 4.6.x environments (the new call form); the pre-4.6 branch is validated by CPU collection/inspection only on this branch.

Draft on the fork to stage the upstream submission; supersedes the exploratory stack (#7/#8/#9)

Review pass (2026-08-30)

• Version compare is now major/minor only: prerelease/local version strings ("4.6.0.dev0", "+git...") of 4.6+ wheels were misclassified as pre-4.6 by the full-triple compare.
• The probe also catches ValueError from unparsable version strings (re-vendored builds), not just PackageNotFoundError.

[wolfcomos]

Closing — decoupled from the four-over-six series (nothing in it is needed by pytorch#4851–pytorch#4853); branch preserved.