fix(cuda): isolate concurrent hardware floor probes

[Qiiks]

Summary

• Release the global probe-cache mutex before running a worker child; per-worker Arc<OnceLock<_>> entries preserve one probe per worker and deliberate success/failure caching.
• Add a regression where a worker confirms it is stalled; another worker must complete before the first is released.
• Document the Linux ELF cuBLASLt runtime dependency requested in Windows owned-CUDA worker: delay-loaded cuBLASLt + sidecar packaging gate #18.

Follow-up to #18 discussion_r4037256258. The separate certification worker-path finding is already addressed in merged master and is not changed here. Based directly on current master; excludes the separate batch-advice change.

Verification

• Old global-lock implementation: regression fails with unrelated probe blocked behind stalled worker: Timeout.
• Fixed implementation, including after rebase: cargo test -p synapse-module --lib cuda_floor_probe -j 1 — 3 passed, 1 explicitly ignored hardware test.
• cargo clippy -p synapse-module --lib -j 1 -- -D warnings and cargo fmt --all -- --check pass.
• No GPU benchmark or deployment change; running native consumers remain untouched.

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}

---

Summary by cubic

Previously, the global CUDA probe-cache lock was held while a worker child ran, so a stalled hardware probe blocked unrelated probes. The lock is now released before running the child; per-worker Arc<OnceLock<_>> entries preserve one probe and deliberate success/failure caching per worker.

• Adds a regression test where a second worker completes while the first is still stalled.
• Documents the Linux ELF cuBLASLt runtime dependency: missing libcublasLt.so.12 makes --probe-floor exit 127 before the driver-only probe.
• No GPU benchmark or deployment change; running native consumers remain untouched.

^{Written for commit 9f9f5ec. Summary will update on new commits.}

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[cubic-dev-ai]

No issues found across 2 files

_{You’re at about 94% of the monthly reviewed-line limit. You may want to disable incremental reviews to conserve quota. Reviews will continue until that limit is exceeded. If you need help avoiding interruptions, please contact contact@cubic.dev.}

_{Re-trigger cubic}

[synapse-alfonso]

Verified beyond the description and merging. The regression test is real: I spliced master's probe body under this branch's tests and it fails exactly as you reported — unrelated probe blocked behind stalled worker: Timeout — then passes with the Arc<OnceLock> entry. Three probe tests green locally, clippy clean, CI green, Cargo.lock untouched. The shape is right: the map lock lives only as long as the entry lookup, the child runs outside it, and per-worker serialisation plus failure caching survive unchanged.

Landing through the CI train now with the Linux DT_NEEDED note you added, which closes the packaging asymmetry from the #17/#18 hardware pass.