cuda.coop: add logical Warp Load and Store

[tpn]

Why this is needed

The preceding stack establishes complete Block and physical-Warp Load/Store
plus topology-aware storage and synchronization. This PR proves that those
same contracts support logical subwarps without changing generic dispatch,
caching, or rewrite orchestration.

Reviewer-visible behavior

• Portable and qualified cuda.coop.load() and cuda.coop.store() accept
  this_warp().group_by(width) with the same call shape, dtype inference, and
  lowercase selector strings.
• Supported logical widths are 1, 2, 4, 8, 16, and 32 lanes. The enclosing
  block must contain complete physical warps.
• direct, striped, vectorize, and transpose map to the corresponding
  CUB WarpLoad and WarpStore algorithms at the selected width.
• Each logical group receives an x-major consecutive tile origin before the
  caller offset, and valid_items is relative to that logical tile.
• Partial transpose loads without oob_default preserve invalid payload slots
  using the logical lane rank.
• direct, striped, and vectorize remain allocation- and barrier-free.
• transpose uses one compiler-owned storage slice per logical group and an
  exact-mask syncwarp, allowing sibling logical groups to diverge.
• Invalid widths, incomplete physical warps, non-string or unknown selectors,
  unsupported selectors, and explicit Warp TempStorage fail before provider
  materialization.

Validation

• The exact head passed 998 source-tree tests with the GPU hidden.
• All 148 focused Warp runtime tests passed on an RTX PRO 6000, including all
  supported widths and algorithms, all ten numeric dtypes, logical tails,
  x-major 3-D blocks, divergent subwarps, and isolated invalid-runtime traps.
• The exact final-stack wheel at 9b299e894f passed 607 contract, packaging,
  public-surface, and activation tests, 50 real NVRTC LTO/nvJitLink compile
  tests, and all 366 runtime tests on the same GPU.
• Changed-file pre-commit and git diff --check passed.

Stack and scope

This draft is stacked on #11205. It supports logical groups formed with
this_warp().group_by(width); mappings such as warps_within_block and
runtime ThreadGroup query or synchronization methods remain out of scope.
CUDA 12/13 L4 and Windows CI remain external readiness gates.

[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.