This is only needed on samd51, and most of the complexity can be
implemented just once instead of in each audio sample type.
This saves 968 bytes flash on feather rp2350. More importantly,
it'll save some fraction as much on samd21 where stereo audio output
is not supported at all (but savings will be less overall because
fewer audiosample types are available)
Other builds besides raspberrypi are broken & raspberrypi is not even
tested.
.. and add a very basic audioeffects test, showing that it plausibly
is working
I had to address several build errors that occurred in the Unix build,
mostly related to conversion from FP types to integral types (replaced
by explicit casts) and by
accidental mixing of regular & f-suffixed floating constants (replaced
with the MICROPY_FLOAT_CONST macro)
Particularly this change could use consideration:
```diff
- self->max_echo_buffer_len = self->sample_rate / 1000.0f * max_delay_ms * (self->channel_count * sizeof(uint16_t)); // bytes
+ self->max_echo_buffer_len = (uint32_t)(self->sample_rate / 1000.0f * max_delay_ms) * (self->channel_count * sizeof(uint16_t)); // bytes
```
The buffer length is being calculated in floating point based on the
millisecond delay & the sample rate. The result could then be a fractional
number such as 529.2 for a 12ms delay at 44.1kHz. Multiplying a floating
number by the size required for each echo buffer item
(`(self->channel_count * sizeof(uint16_t))`) could yield a number of bytes
that doesn't correspond to an integral number of buffer items. I grouped
the float->int conversion so that it converts the number of echo buffer
items to an integer and then multiplies by the size of the item.
BlockBiquad takes kind, f0 (center frequency) & Q (sharpness)
block type arguments and calculates the actual filter coefficients
every frame.
This allows the filter characteristics f0 and Q to be changed dynamically
from LFOs & arithmetic blocks.
A new manual test demonstrates this on a host computer, playing a simple
tone that is dynamically filtered.
