[bdbarnett]

Since the sampler gained its low-pass filter, modulation envelope and LFOs (#337), every voice slot carried their state whether a note used them or not. A voice grew from 112 to 256 bytes, and a sampler allocates two slots per voice (the voice and the tail that fades its last note) in one block. Sampler(voice_count=32) went from 16 KB to 27 KB, and its voice table from 7 KB to 16 KB.

That broke SoundFont instruments on a crowded heap. In an 8 MB MicroPython heap, once a General MIDI bank (FluidR3_GM.sf2) has loaded, the 16 KB table no longer fits in any free block. The instrument raises MemoryError: memory allocation failed, allocating 16384 bytes, and its notes are silent.

What changes

The filter, envelope and LFO state moves out of the voice into its own record, audiodsp_sampler_mod_t. A voice gets one pair of records (its own and its tail's, 272 bytes on a 64-bit build) the first time it needs them: a play() that sets any of those arguments away from its default, a cutoff set on the voice, or a nonzero vibrato_depth. The voice keeps the pair after that. A voice that never asks carries none, and a sampler whose voices all ask allocates 32 small pieces rather than one large one.

The voice's byte-sized fields (stage, loop_mode, looping, mods) now share the padding before position. A plain voice is 112 bytes on a 32-bit board, the same as before #337, and 120 on a 64-bit desktop. A size check in audiodsp_sampler.c fails the build if the voice grows past that.

Measured on the unix port (64-bit):

	v0.6.11	this PR	before #337
voice slot	256 B	120 B (112 B on 32-bit)	112 B
Sampler(voice_count=32), stereo	27328 B	18624 B	16064 B
Sampler(voice_count=32), mono	25696 B	16992 B
Sampler(voice_count=128), stereo	99424 B	64608 B
records once all 32 voices use a filter	(included)	7168 B, in 32 blocks

Renders don't change

A record attached to a voice that is already sounding starts as play() would have left it for a note that asked for none of this. The note goes on rendering exactly as it would have.

• tests/parity/sampler_probe.py gives byte-identical output before and after the change on CPython and on the unix MicroPython build, and the two agree. Its closed section turns a filter on live, on a voice that had played plainly.
• A second script covers paths the probe doesn't. It hands modulated and plain notes to their tails, plays a plain note on a voice that already has a record, attaches a record mid-note through vibrato_depth, and runs mono and stereo. Its output is byte-identical before and after on both runtimes. Two planted faults each changed its output: not copying the record to the tail, and a wrong default in a new record.
• tests/test_cpython_audiosampler.py, tests/test_cpython_audiosampler_modulation.py and tests/test_binding_parity.py pass.
• The CircuitPython bindings compile with -Wall -Werror against CircuitPython 11's unix headers. No CircuitPython build was run.

With this change alone, the mpvst unix engine at an 8 MB heap loads FluidR3_GM.sf2 and plays: SoundFont note 60 peaks at 11094 and SoundFont Drums note 36 at 13222. Forty notes held at once leave 2 MB free.