Add relocatable CUTLASS provider AOT packs

[tpn]

Why this is needed

Some deployments need to compile CUTLASS providers once and reuse the exact
native inputs in later processes. This final layer adds an explicit,
relocatable AOT pack workflow without changing primitive calls or weakening
exact cache identity.

python -m cuda.coop._aot_cli capture --output providers.coop-aot -- \
  python my_kernel.py
python -m cuda.coop._aot_cli run --pack providers.coop-aot \
  --mode required -- python my_kernel.py

What this PR includes

• Python capture/use/inspect API and source CLI
• canonical manifests, atomic publication, relocation, and integrity checks
• exact identity and symbol matching before provider resolution
• required, auto, and off behavior with final-link integration

Stack

7 of 7, based on PR #11048.
This is the top of the cuda.coop stack.

Local validation

• 85 focused AOT API, manifest, resolver, finalizer, and public-surface tests
  passed
• captured, inspected, and relocated a 244,388-byte radix/TopK provider pack
• a fresh required-mode exact hit passed final linking and GPU output without
  creating a JIT cache artifact
• auto exact miss and off mode each compiled one JIT artifact and passed GPU
  output
• source imports, compileall, changed-file pre-commit, Ruff, import audit,
  and git diff --check passed

Packaging tests, wheel 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.