Also skip the first loss values being generated since they're
biased towards "not lost" due to the initialization.

Fails ubsan because memcpy declares args as non-null

No longer needed now that PLC is trained with PyTorch stack

Should handle the history in a more consistent way.
Slightly increase the model size and re-enable biased band loss in
training.

mostly untested

Should match the TF2 code, but mostly untested

Also default to disabling dnn float debugging

Still don't quite know what I'm doing

Compilers sometimes replace vmlaq*() with fmul+fadd instead of fmla.
Trying to use vfmaq*() instead when possible.

Finished adversarial training on 800k model. Also, move weights to a new
location.

Adds RTCD tables for compute_activation() and compute_conv2d()

Avoids having to write intrinsics for simple loops

Still missing some intrinsics

Used for DNN matrix multiplies