fix(audio): replace queue clears with bounded callback buffering

[sunshinewithmoonlight]

Summary

• Replace SDL queued-audio playback and full-queue clears with a bounded PCM ring.
• Start playback after a small prefill and fill callback underruns with silence.
• Drop only the oldest complete 10 ms blocks when the high watermark is exceeded.
• Add low-frequency diagnostics across the SDL callback and ring watermarks.

Problem

• Current AudioManager requests 1024 samples, accumulates SDL queued audio,
  and clears the whole queue above 16000 bytes.
• Full clears can cause audible gaps, and the retained queue adds latency.

Scope

• Akira client playback only.
• No video, input, network protocol, codec, or release-version changes.
• Server-side packet redundancy and scheduling fixes live in
  OrbisWindowBridge and are not part of this PR.
• The library/chiaki-ng submodule is not modified.

Follow-up from hardware testing

• Initial hardware testing exposed audible distortion and noise that the
  original buffering change did not resolve.
• Use an absolute discard sequence so high-watermark drops always advance the
  effective read position.
• Preserve PCM during prefill instead of consuming it while outputting silence.
• Remove the 1.80x playback gain.
• Serialize producer writes with the SDL callback PCM access.
• Rate-limit summary logging correctly.

Tests

• Host ring-buffer unit tests cover prefill, wrapping, high-watermark drops,
  underruns, clear/new-session behavior, non-stereo parameters, and concurrent
  producer/consumer stress.
• make test: 174 passed, 0 failed.
• Focused ASan/UBSan and TSan runs: 10 passed, 0 failed each.
• Akira build/CI result: pending.
• Hardware retest of the rebuilt client is still pending.

Risk and rollback

• Real-time callback synchronization and session cleanup are the main risks.
• Revert the focused commit to restore the previous AudioManager.

This client-side change is not claimed to remove the full end-to-end latency by
itself; the sender packet history and pacing fixes are separate.

[xlanor]

I currently have a bunch of fixes in flight around something similar, so will only look at this after.

I ask that before sending any PRs you please test this on hardware first as well