harden free-threaded lock ownership

[Qubitium]

Summary

• require constructor-owned locks for Looper cache state, module wrappers, thread-pool telemetry, and patched autotuners
• pre-create locks for the known extension registry and create workspace locks once per previously unseen device, with an allocation-free established-key path
• drain every submitted worker future before propagating any BaseException
• use built-in dictionary equality while retaining synchronized mutation and iteration snapshots
• retain best-effort device-handle shutdown while logging ignored close failures at DEBUG level

Why

Several fallback paths dynamically created locks during ordinary API calls or allocated a temporary lock on every workspace lease. That adds overhead and makes free-threaded ownership harder to reason about. Worker collection also used an exception allow-list, allowing a custom BaseException subclass to bypass the drain guarantee while other workers could still be using shared subset state.

The locks remain access-scoped or key-scoped. Worker tasks are still submitted before result collection, so module processing remains parallel.

Validation

• focused free-threaded lock and concurrency suite — 29 passed
• coverage-instrumented targeted suite — 41 passed, 1 skipped
• CUDA workspace contention test — 1 passed
• changed executable code — 36/36 lines and 8/8 branches covered (100%)
• Ruff checks on the modified sources and tests — passed
• repository source audit — 0 lazy lock-attribute checks, 0 setdefault(..., Lock()) factories

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[devin-ai-integration]

Devin Review found 1 new potential issue.