Limits on attack time for lookup table oscillators

[dpwe]

Lookup table oscillator (sine, pulse, saw, etc.) interpolate their amplitudes over adjacent frames, to avoid clicks when amplitudes change rapidly. This means the fastest that a note can start from zero is 5.8 ms (one frame), which is a little too long for the attack of some percussive instruments, like a punchy piano. It would be nice to be able to go down to ~1 ms.

Using shorter frames (e.g. 128 samples instead of 256) would improve this, at the cost of greater processing overhead.

We could also special-case attack frames, perhaps with a split linear amplitude fade (ramp up for some sub-portion of the frame, then have a different slope for the remainder).

[dpwe]

I tried setting AMY_BLOCK_SIZE to 64, and everything worked pretty well (except for PARTIALS, see #81). 64 gives us a minimum attack time of under 1.5 ms which is pretty good, but it means the per-block processing overhead (which involves floating point operations etc) is 4x greater, which will reduce the total number of oscillators we can manage for a given CPU capacity.

See branch test_short_blocks

[dpwe]

This was mostly solved by a06dad4 which shifted the amplitude envelope one block earlier. Thus, if the envelope generator value at time 0 is 1.0, the osc amplitude starts at 1.0 (previously it would fade from 0 to 1 over the first 5.8ms). This makes possible instantaneous attacks (attack time=0, good for percussion) but still leaves the second fastest attack time limited to 5.8ms.