cudnnAddTensor() function and updated the specs and asserts accordingly.

kernel was still running.

CUDA version of add_bias_gradient().

batch_normalize_conv.

of add to them so that it's consistent with how the layer interface expects
this to be done.

for this object.

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.