Including removing useless code

Removing one of the 2d conv layers for pitch estimation reduces
complexity without noticeable degradation. FARGAN model has more
adversarial training.
Also, no need for the double precision in the low-pass filter.

Gets us about 2x speedup on x86

Shift computation to concealment

And references to nnet_data.h

Those intrinsics don't actually require alignment so we're OK