See https://github.com/mozilla/LPCNet/commit/bffdcee95#commitcomment-46372726

When implementing using SSSE3 or AVX2, our dot products can saturate
if two adjacent weights sum to more than 127.

Not sure why CuDNNGRU doesn't get used by default, but we need
to explicitly use it to get things to run fast.

was increased too much in 713d53e8a

Reciprocal approximation could cause the sigmoid output to be
greater than 1.0.

Would cause unused pitch embedding vectors to be used

No change otherwise

preventing the pitch from going above 255

According to David Rowe, when rand() returns RAND_MAX (which is likely
for 16-bit output), we end up producing a click.