tensors.

train_one_step() member function.  Also improved how the host to device transfers
are overlapped with kernel computation.

to finish before overwriting the device memory with updated values from the
host.

currently active device id if the user changes the active device via a call to
cudaSetDevice().

than adding them.  This way, the gradient buffer can be used as scratch space
during the loss computation.

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

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