runs on the GPU, and made affine_transform() take only tensors.

use either CPU or GPU.  Fixed a bug in gemm().

form.

in-place.

code based on how dlib was built.  All the layer implementations will interact
with these functions.

things up a bit.