fix(audio): count complete frames in JS Audio.write

[toprameen]

JS Audio.write counts each unmatched stereo sample as half a frame, so two separate one-sample writes can report one queued frame. Count complete frames per write, matching native behavior, and preserve whole-request rejection above 8192 samples.

Add a focused regression that prints 1 before the fix and 0 afterward through source/compiled CLIs and emitted Bun/Node. Local Audio fixtures and capacity checks pass; repository gate 54/54.

[Lorenzobattistela]

Thanks! The fix reproduces: on main, JS prints 1 for two one-sample writes, and on this branch it prints 0, matching native. It also fixes a second case: on main, one odd sample followed by a full 8192-sample write is rejected in JS (queue 0.5 + 4096 > 4096), while native accepts it and reports 4096. All gates are green merged onto main 24a997f.

The JS device still doesn't match native, though. A few things before merging:

1. The queue still drains in fractions of a frame. After filling the ring (4096 frames) and draining 0.3 of a frame by the clock, queued is 4095.7 and the write reports 4095. A caller that sizes its next write as 4096 - reported (the Sound.write loop in demos/app_slash_boss_3d does this) sends 1 frame, which is refused because 4095.7 + 1 > 4096. Native accepts it: its queue is written - read, both whole frame counts. A Node harness with a fake clock reproduces it on this branch.

2. Because of (1), I'd fix the cause rather than round the input. Keep queued a whole number of frames, like native's counters: drain only whole frames by the clock, and carry the leftover time in at instead of dropping it into queued. Then Math.floor(n / 2) is the only rounding left, and the result no longer needs Math.floor(audio.queued).

3. Please add a C twin for pause (tests/io/audio_whole_frames.c, returning the handle unchanged). Without it the native build fails with "no .c import: pause", so the gate drops the C lane and the whole-frames rule is never checked on native. With an identity pause the test prints 0 natively too. Then the "JS only" line in the header can go.

Smaller:

• n > 8192 ? 0 changes nothing: for n > 8192, frames > 4096 already fails the room check. Dropping it removes the bare 8192, and the 4096 that's left should be one named constant next to audio.c's IO_RING.
• pause works by writing the handle's private rate field. If audio_open.js ever renames it, pause does nothing, and the old n/2 bug still prints 0 (0.5 - ε + 0.5 rounds down), so the test would stop catching the regression. A short comment in audio_open.js naming the test that depends on rate would help.