This adapts the "inline assembler" code from the UF2 bootloader, which
in turn is said to be adapted from the arduino neopixel library.
This requires the cache remain ON when using M0, and be turned OFF on M4
(determined by trial and error)
Testing performed on a Metro M4:
* measured timings using o'scope and found all values within
datasheet tolerance.
* Drove a string of 96 neopixels without visible glitches
* on-board neopixel worked
Testing performed on a Circuit Playground Express (M0):
* Color wheel code works on built-in neopixels
* Color wheel code works on 96 neopixel strip
As a bonus, this may have freed up a bit of flash on M0 targets. (2988 ->
3068 bytes free on Trinket M0)
Closes: #2297
1905d0746d