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