fix(sound): play cues in full stereo from the first sample

[Nick6489]

Problem

Stereo sound cues played in mono, most audibly the connect sound.

Every cue was played through a rodio Sink. Sink wraps its playback queue in a
channel converter the instant it is built, while that queue is still empty, and an
empty rodio queue reports one channel (its Empty placeholder source). The
converter is therefore locked to mono to stereo for its first batch: the opening
512 samples (256 frames, ~11 ms) of whatever plays next get read as mono, with
left and right interleaved into one stream and each sample duplicated to both
speakers, which also time-stretches that slice. Only after that batch does the
converter re-read the source and discover it is actually stereo.

The default cues are short and their stereo image sits in the opening transient, so
connect/disconnect came out sounding mono while longer, noisier cues masked it.

Reproduced by driving the real Sink -> queue -> mixer wiring with the default
connect_success.ogg: 512 leading output frames with L == R exactly, versus 0
for the bare decoder.

Fix

Decode each cue fully into a SamplesBuffer and hand it straight to
OutputStreamHandle::play_raw instead of going through a Sink.

• A SamplesBuffer carries its real channel count from the first sample, so
  nothing is folded. The reproduction now shows 0 mono frames and
  stereo-difference energy identical to decoding the file directly.
• A non-48 kHz output device now resamples the whole buffer in one correctly
  configured pass, instead of the old path's per-packet re-bootstrapping.
• Sound lifetime for the exit wait is tracked with an AtomicUsize that rodio's
  Done decrements when a cue ends, replacing the Vec<Sink> + sink.empty()
  poll. Volume is applied by scaling samples during decode.

No public API change: play_sound_file, wait_for_playback, SoundPlayer, and
EXIT_SOUND_TIMEOUT keep their signatures.

Tests

decode_to_buffer is split out as a device-free helper so the pipeline is testable
without an audio device:

• a_stereo_sound_is_buffered_without_being_folded_to_mono -- the regression
  guard: the decoded buffer still reports 2 channels and keeps distinct L/R
  samples.
• volume_scales_the_buffered_samples
• a_zero_length_sound_is_not_played

cargo fmt --all --check, cargo clippy --workspace --all-targets -- -D warnings,
and cargo test --workspace all pass.

🤖 Generated with Claude Code