fix(realtime): avoid false keepalive timeouts under load

[LauraGPT]

Summary

• keep server WebSocket pings enabled every 20 seconds, but disable ping-timeout closure by default
• make positive ping timeouts an explicit operator choice above measured worst-case decode/queue delay
• add benchmark controls and JSONL fields for client ping interval/timeout
• document the keepalive boundary in English and Chinese, including why ws-max-queue doesn't fix control-frame timeouts

Fixes #3528.

Root cause and controlled reproduction

Under concurrent long-audio load, synchronous VAD work and queued model decoding can delay WebSocket control-frame handling beyond 20 seconds even while a session is making valid progress. The previous ping_interval=20, ping_timeout=20 default then closed healthy connections with code 1011.

On one H100 80 GB service, using the same paced 47-second Chinese WAV and 20 concurrent clients:

Server setting	Finals	Errors	Elapsed	STOP p95
v1.4.4 default, timeout 20s	4/20	16 (keepalive ping timeout)	65.467s	18.429s among survivors
v1.4.4 explicit timeout off	20/20	0	81.879s	34.838s
this PR, default options	20/20	0	67.021s	19.970s

The timeout-off control proves that ws-max-queue isn't the failure boundary: it limits incoming data messages, while ping/pong are WebSocket control frames. The exact latency varies between runs, so the pass/fail and close-code result is the relevant comparison.

The reporter independently observed the same behavior on L20: 47-second/16-client sessions complete when server ping timeout and client pings are disabled. This PR's hardware result is H100 evidence, not a substitute for their L20 measurement.

Behavior and tradeoff

The default remains ping_interval=20, so healthy connections still exchange keepalive traffic. ping_timeout=None prevents compute backpressure from being mistaken for a dead peer. Deployments that require server-side pong expiry can set a positive --ws-ping-timeout after measuring their worst-case production queue/decode delay and coordinating with gateway idle timeouts.

Verification

• 113 passed across realtime service, benchmark, install docs, vLLM source docs, and growth-plan tests
• post-fix H100 default-options replay: 20 final messages, 0 client errors, no server traceback/error
• git diff --check
• Python bytecode compilation for changed Python modules and tests
• signed commit with DCO trailer