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.

The sinc filter offset was incorrectly handled. Since it perfectly compensates
for the analysis offset, nothing has to be done.
Also, the preemphasis memory was never initialized.

Gets us about 2x speedup on x86

Controls whether deep PLC is enabled

Shift computation to concealment

No dependency on the data files if no DNN code enabled

And references to nnet_data.h

Signed-off-by: Jean-Marc Valin <jmvalin@amazon.com>

Signed-off-by: Jean-Marc Valin <jmvalin@amazon.com>