Add cuda.coop Exchange and Shuffle with Numba-CUDA-MLIR

[tpn]

Why this is needed

Kernels often need to rearrange per-thread values after loading them. This
adds common Exchange and Shuffle APIs with their Numba-CUDA-MLIR providers,
using the group, storage, and synchronization contracts from #11200.

What changes

• Common exchange converts blocked and striped layouts for Block, physical
  Warp, and logical Warp groups. The backend API also supports Block-only
  warp-striped, scatter, guarded/flagged, ranks/flags, and warp-time-slicing
  forms.
• Common Block shuffle provides unit up and down operations. The backend
  API adds scalar offset and rotate forms.
• Both operations return fresh payloads and preserve inputs, including ranks
  that CUB uses internally. Mode selectors are lowercase strings.
• Storage-consuming operations use per-group scratch and the corresponding
  Block or exact-mask Warp reuse barrier. Payload shapes, dtypes, ranks,
  flags, selectors, and runtime distance ABIs are checked before compilation.

blocked = coop.exchange(
    coop.this_block(), striped, mode="striped_to_blocked"
)
shifted = coop.shuffle(coop.this_block(), blocked, mode="down")

Compile tests verify barrier placement and masks at all six Warp widths.
Runtime tests repeatedly reuse scratch while sibling groups execute different
iteration counts, with each group converged. Opt-in racecheck tests compare
production barriers with a test-only control that removes the reuse barrier.

Warp scatter and Shuffle boundary-output projections remain unsupported.
Invalid runtime rotate bounds or offset values outside the signed-32-bit
provider ABI trap the device; failure tests use isolated processes.

Stacked on #11200.

Validation

Fresh installed-wheel checks at 0ce85aab9385:

• Host contracts and packaging (full suites): 516 passed on Python 3.14.
• Numba backend unit/compiler tests (focused, GPUs hidden): 92 passed on Python 3.14.
• GPU runtime tests (focused): 11 passed on Python 3.14.

Wheel validation and source, installed-package, and bundled-header comparisons passed.

Remote CI passed at 0ce85aab9385: GitHub Actions run. All exact-head checks and statuses, including pre-commit, are successful or conditionally skipped.

[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.

[tpn]

Regarding the three suppressed comments in this review:

• Warp scatter is outside this API's supported surface. Physical and logical Warp groups accept only blocked_to_striped and striped_to_blocked; _mode_token() rejects scatter before provider selection. The WarpGroup overload also requires ranks=None, and the existing physical/logical rejection test checks this boundary. The lowerer's rank-copy branch does not make Warp scatter publicly supported, so no runtime case or API expansion is needed here.
• Added the Exchange/Shuffle implementation modules, stubs, and Block Exchange, Block Shuffle, and Warp Exchange headers to the standalone wheel validator. The wheel build now checks these files directly.
• Kept scalar Shuffle dtype inference unchanged. The frontend passes an ir.Var; GroupPlanningContext.dtype() follows its definitions, and _dtype_definition() handles ir.Const through _python_scalar_dtype. NumPy scalar constructors such as np.int32(4) use the existing scalar-cast path. The claimed missing constant-dtype path is therefore already implemented.

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 19m 17s: Pass: 100%/8 | Total: 1h 03m | Max: 18m 40s

See results here.