feat(sounds): crash-isolated sound server and dual-channel output modes

[larsrollik]

What

Make sound output robust against a native sound-device fault, and fix the stereo output so playback doesn't depend on which conductor is wired.

Crash-isolated sound server

SoundServerClient is a drop-in for StereoSound that runs PortAudio in a spawned subprocess. A native segfault or hang in the sound device now takes down only the child; the client detects a dead worker (proc.is_alive()) and degrades to no-sound instead of propagating the fault. (A native segfault can't be caught in-process, so isolation is the only real defense.)

Latency design: spawn start (Windows-safe), a blocking Queue.get() in the server (event-driven, avoids the ~15.6 ms Windows timer granularity a poll loop would hit), timeBeginPeriod(1) + elevated priority on Windows, best-effort nice on Unix, pre-registered cues so a trigger is a fire-and-forget integer put.

Channel modes on StereoSound

• channel_mode="both" (default) — the tone on both output channels, so playback no longer depends on which L/R conductor is wired.
• channel_mode="ttl" — the tone on one channel plus a full-scale sync level on ttl_channel, a hardware onset marker for a BNC/DAQ input.

Measurements

Added trigger latency (client put → server get, the isolation overhead), Linux:

condition	p50	p99	max
idle	0.03 ms	0.05 ms	0.18 ms
all cores at 100%	0.03 ms	1.43 ms	3.66 ms

End-to-end on real Bpod hardware via a BNC loopback (softcode → handler → audio → BNC1High, timed on Bpod's own clock):

backend	p50	p99
in-process (today)	19.75 ms	21.55 ms
server (isolated)	20.70 ms	21.35 ms

The isolated server is within run-to-run noise of the in-process path. The ~20 ms absolute is audio output buffering (pipewire quantum + DAC + softcode USB round-trip), shared by both paths, not introduced by the subprocess.

Channel-mode check on hardware: both 50/50 hits; ttl fires ~7 ms earlier (marks true buffer onset rather than waiting for the sine to climb its fade-in) and is 30/30 at realistic trial spacing.

Verification

• scripts/bench_sound_server.py — IPC latency benchmark (run on Windows to confirm parity: uv run python scripts/bench_sound_server.py -n 5000).
• scripts/measure_sound_latency_bpod.py — Bpod BNC-loopback end-to-end harness.
• Unit tests for both channel modes (tests/test_sounds.py).

Follow-ups (not in this PR)

• Windows IPC p99 run to confirm cross-platform parity.
• Switching the tasks over to SoundServerClient (this PR only adds the capability).