Add the CUTLASS backend for cuda.coop

[tpn]

Why this is needed

The portable cuda.coop layer defines cooperative operations, but it does not
yet make them usable from CUTLASS DSL kernels. This stack layer adds the
CUTLASS compiler substrate and implements the initial load/store, reduce, and
scan cohort behind the same API introduced in #11035.

import cuda.coop.cutlass as coop

items = coop.load(coop.this_block(), source, coop.ThreadData(2))
total = coop.sum(coop.this_block(), items)
coop.store(coop.this_block(), destination, items)

What this PR includes

• CUTLASS DSL 4.8 activation and automatic backend registration
• lowering and provider generation for load/store, reduce/sum, and scan
• provider compilation, final linking, scratch storage, and CUB/CUDAX routing
• semantic, header, toolkit, and compiler cache identity with integrity and
  fork-safe locking

PCH and AOT land in stack layers 6 and 7.

Stack

2 of 7, based on #11035. The next layer is the Numba-CUDA-MLIR backend.

Local validation

• 552 focused CUTLASS source tests passed across unit, provider compilation,
  linking, and GPU runtime coverage
• 101 affected import/boundary tests passed after the final stack cleanup
• source-root and root-free final-link GPU lanes passed in the local CUTLASS
  DSL 4.8 environment
• source imports, compileall, changed-file pre-commit, and
  git diff --check passed

Wheel and installed-package qualification, GitHub CI, and automated AI review
are deferred for now.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.