fix(realtime): decouple bounded receive buffering from inference

[LauraGPT]

Summary

Related to #3528; this PR does not close the issue or claim the reporter's L20 workload is fixed.

• Receive independently of session inference into a bounded, ordered application FIFO (128 messages / 16 MiB by default). Overflow explicitly closes with 1013; it is not successful final delivery.
• Defer only provisional decodes when the next queued message is audio. Preserve PCM and command order, completed-segment inputs, and COMMIT/STOP paths.
• Join owned receive/worker tasks on exit and suppress successful results after an in-flight operation outlives receive failure.
• Document the queue/ping relationship, memory boundaries, source-vs-PyPI availability, and hardware acceptance requirements.

Verification

• Exact candidate on ind-gpu8, Python 3.12.3: 142 tests passed with websockets 17.1; 142 passed with 10.4.
• New real-loopback tests use synthetic acoustic sessions: finite blocked-decoder bursts retain Pong and ordered audio/final delivery; sustained overflow gives 1013; connections have independent queues.
• Earlier candidates fail the added receive/cleanup guards. The payload-lifetime test still rejects persistent references after one event-loop completion turn.
• Repository docs: 268 passed / 4 dependency-related skips, followed by all 23 tests in the affected browser/Sphinx modules passing without skips. All four missing checks were executed.
• Two site builds: 267 files / 202 HTML pages each, validation passed and all bytes identical.
• Independent review of source, workflow and docs: no remaining findings. Test authorship and review are distinguished in the retained evidence.

Remaining Acceptance

Draft intentionally: the new Python 3.10/10.4 and Python 3.12/17.1 CI must pass on this exact head. Actual model low-concurrency behavior, final transcript quality, preview freshness, shared-engine fairness and the reporter's paced 47-second / 16-client L20 workload still require acceptance. H100 availability is not L20 evidence.

Fewer partials may change context/fallback observations. No transcript-equivalence or throughput improvement is claimed. Queue limits bound application payloads, not all protocol/session/process memory. Waiting for thread completion does not force-stop a hung inference backend.

No PyPI release, production deployment, issue closure or repeated request for the reporter's parameter grid is included.