Add cuda.coop data collectives

[tpn]

Why this is needed

The walking cohort establishes the shared compiler and API contracts. This
layer expands that contract vertically with the data-movement and
segmentation collectives needed by real cooperative kernels, keeping the
portable surface and both compiler backends synchronized.

What this PR includes

• exchange and shuffle
• adjacent difference and discontinuity
• histogram and run-length decode
• matching portable contracts, CUTLASS lowering/providers, and
  Numba-CUDA-MLIR planning/lowering

Ordering and selection remain isolated in the next stack layer.

Stack

4 of 7, based on PR #11045.
The next layer adds Merge Sort, radix operations, and TopK.

Local validation

• 227 portable contract tests passed
• 37 focused CUTLASS 4.8 tests passed
• the complete Numba source tree at this stack point passed 242 unit, compile,
  and GPU-runtime tests, including complex128 value callbacks
• source imports, compileall, changed-file pre-commit, Ruff, import audit,
  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.