Add a General MIDI synthesizer: tecs.audio.synth over rustysynth

[albanmgicquel]

Adds tecs.audio.synth: General MIDI from a SoundFont (SF2), rendered in-process by rustysynth (MIT, no dependencies) as one voice of the existing mixer, so group gain/mute/pause, master gain and maxVoices apply unchanged.

const synth = require("tecs.audio.synth")
const font, err = synth.loadSoundFont(audio, path)
const s, err = synth.new(audio, font, {group = "music", gain = 1.0, polyphony = 64, effects = true})
s:send(status, data1, data2, at?)       -- plus noteOn / noteOff / controlChange / programChange / pitchBend
s:allNotesOff(channel?, immediate?); s:reset(); s:now(); s:setGain(g); s:destroy()

• Scheduling: at is a time on the audio clock (now(), frames rendered over the rate). Events land on rustysynth's 8-frame block grid at or after at (never early, at most ~0.15 ms late at 48 kHz), so a given schedule renders identically every time; a nil or past at means the next boundary, in send order. Block 8 measured at 3.3% of one core with FluidR3, 48 notes held, effects on (1.9% at rustysynth's default 64).
• Real-time: a bounded lock-free SPSC ring (4096) feeds a preallocated schedule (4096); no allocation or blocking in the callback. A full queue refuses the send: the send calls return false (and the drop is counted) rather than blocking or growing.
• Headless: with no output, or the silent backend, the synth gets a silent port (every call succeeds, now() stays 0); a test backend can supply openSynth to log events.
• Extension point: a hardware MIDI output (e.g. midir: CoreMIDI, WinMM, ALSA) would be a second SynthPort with the same send/at API; documented, not implemented.
• Also: tecs.audio moves to audio/init.nupp (to host the child module; require("tecs.audio") is unchanged); the backend contract gains optional synth fields; a README design record explains why General MIDI is synthesized in-process (desktop OS GM synths are software anyway; one sound on every platform; mixer integration; determinism). No SoundFont ships; THIRD_PARTY_NOTICES says a game ships its own font's licence.

Tests: Rust scheduling/queue/reset/clock tests with a recording synth; SoundFont-gated rendering tests (TECS_TEST_SOUNDFONT): a note scheduled at frame 4800 renders the same samples as one at 0 shifted by 4800, identical renders for the same schedule, gain, immediate cut vs release, reset discarding a scheduled note, pause freezing the clock; a Nupp suite (audiosynthtest) over a logging backend. cargo test -p tecs-audio, clippy -D warnings, cargo fmt --check, nupp check --strict, nupp task format-check and nupp test pass. Used in a game with FluidR3_GM, music verified by ear.

Not run here: nupp task docs-check (this toolchain lacks the lunamark rock) and the packaging step. Known: rustysynth's SF2 parser is not hardened against a crafted file (the docs say to load only fonts the game ships); in one early device run the clock appeared to stop at 2.24 s with no stream failure flagged, not reproduced in later runs.