[bdbarnett]

audioinstruments.create("soundfont", ...) and create("soundfont_drums", ...) opened and parsed the bank before making the instrument's own voices. A sampler's voices are one block of memory, and after a large bank (FluidR3_GM, for one) is parsed the heap can have megabytes free in small pieces with no gap big enough. On an 8 MB MicroPython heap with the current audiodsp sampler, whose voices are larger, create() raised MemoryError allocating 16384 bytes and the instrument never played.

Both instruments now make their own pool before opening the bank, and free it again if the bank won't open. A pool you pass in as voice_pool is untouched either way.

On an 8 MB heap with FluidR3_GM, each instrument created alone now plays where it raised before. A second instrument with its own pool, created after a large bank is already open in the same interpreter, can still hit the same wall; make one Voices pool first and pass it to both, as the Voices docstring shows.

Tests: VoicesBeforeBankTest in tests/test_cpython_soundfont.py checks the order for both instruments and that an own pool is freed when the bank fails to open; three of its four tests fail without the change.