cuda.coop: add Numba Exchange and Shuffle

[tpn]

Why this is needed

The preceding stack extends Load and Store through block, physical-Warp, and
logical-Warp providers. This PR adds Exchange and Shuffle as independently
registered primitive families, exercising the provider, result, storage, and
synchronization seams without changing generic activation, dispatch, caching,
or phase orchestration.

Reviewer-visible behavior

• Adds portable exchange for block, physical-Warp, and logical-Warp
  striped/blocked layout conversion.
• The qualified API preserves that shared surface and adds Block-only
  warp-striped, scatter, guarded/flagged, ranks/flags, and warp-time-slicing
  extensions.
• Warp Exchange exposes the shared striped/blocked modes; scatter is not
  exposed for Warp groups.
• Adds portable Block shuffle for unit up and down operations.
• The qualified API preserves that shared surface and additionally supports
  scalar offset and rotate forms.
• All public mode selectors are lowercase strings.
• Exchange and Shuffle return fresh payloads and preserve caller inputs,
  including ranks that CUB may modify internally.
• Storage-bearing forms use implementation-owned leading-pointer storage with
  block or exact-mask Warp reuse synchronization and disjoint slices per
  logical group.
• Payload dtypes and extents, signed ranks, integral flags, selectors, and
  runtime distance ABIs are validated before provider specialization.

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

Validation

• The exact head passed 168 focused public-surface and Exchange/Shuffle tests.
• The exact final-stack wheel at 9b299e894f passed 607 contract, packaging,
  public-surface, and activation tests outside the source tree.
• That wheel passed 50 real NVRTC LTO/nvJitLink compile tests and all 366
  runtime tests on an RTX PRO 6000.
• Changed-file pre-commit and git diff --check passed.

Stack and scope

This draft is stacked on #11207. Shuffle boundary-output projections, Reduce,
Scan, callbacks, and Python operator compilation remain deferred.

Invalid runtime rotate bounds and runtime offset values outside the
signed-32-bit provider ABI trap the device and invalidate that CUDA context;
the implementation tests these paths in isolated processes.

[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.