Add the portable cuda.coop foundation

[tpn]

Why this is needed

This is the first layer of a stacked cuda.coop series. It gives compiler
integrations one backend-neutral contract for CUDA thread groups, fixed-size
per-thread values, temporary storage, and the initial load/store, reduce, and
scan operations.

Keeping these contracts in the root package lets later CUTLASS and
Numba-CUDA-MLIR PRs implement the same calls without coupling the public API
to either compiler.

import numpy as np
from cuda import coop

block = coop.this_block()
items = coop.ThreadData(2, dtype=np.int32)
loaded = coop.load(block, source, items, valid_items=count, oob_default=0)
total = coop.sum(block, loaded)
coop.store(block, destination, loaded, valid_items=count)

What this PR includes

• the cuda-coop package foundation and backend-independent imports
• thread groups, launch facts, argument bindings, ThreadData, and
  TempStorage
• portable load/store, reduce/sum, and inclusive/exclusive scan contracts
• bundled-header discovery and compiler-free contract tests

CUTLASS activation starts in the next PR. Numba-CUDA-MLIR and the remaining
primitive families follow in later stack layers.

Stack

1. This PR: portable core, based on main
2. PR #11044: CUTLASS walking cohort
3. PR #11045: Numba-CUDA-MLIR walking cohort
4. PR #11046: data collectives
5. PR #11047: ordering and selection
6. PR #11048: CUTLASS PCH
7. PR #11049: CUTLASS AOT

Each PR targets the branch immediately below it; all seven are drafts.

Local validation

• 379 source-tree core and header contract tests passed; two packaging-only
  tests were deselected
• source imports and compileall passed
• changed-file pre-commit and git diff --check passed
• CodeRabbit and both required RoboRev reviewers completed for this first
  layer; all accepted fixes are folded into the commit

Wheel builds, installed-wheel qualification, GitHub CI, and backend
compilation are deferred to their owning stack layers. Automated AI review of
layers 2-7 is 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.

[jrhemstad]

16,000 lines is still too large for anyone to meaningfully review. We're going to need to break this smaller. I recommend focusing the first PR on enabling a single algorithm.