Special casing floats decreases runtime to about 50% (applying a 117-tap
filter to 512 points of data goes from 70ms to 32ms)
The top_n/bot_n calculations already meant that the a/c indices were
never out of range. This decreases runtime further to about 15% of
original (11ms)
Timings done on an Adafruit Clue (nrf52840 at 64MHz)
It does of course increase code size somewhat.
7846b0c469