Combine matmuls for GRU gates

[robertknight]

Apply a change similar to 2f0327c by combining the separate matrix multiplications for the 3 gates into a single matmul. This reduces the total number of matmuls from num_directions * seq_length * 6 to num_directions * seq_length * 2.

Using only two matmuls is only possible when linear_before_reset=true, as is the case all models I have worked with so far, since it is what PyTorch and cuDNN support and what TensorFlow uses by default. With the GRU formulation in the original paper, an additional matmul is required since the reset gate has to be computed before the hidden_state @ hidden_weights[hidden_gate] matmul can be done.

Fixes #85

TODO:

• Fix efficiency of binary ops for non-contiguous inputs. This affects accumulating / scaling subsets of gates.
• Fix efficiency of parallel unary ops for non-contiguous inputs. This affects applying activations to subsets of gates. Update: This mostly affected tanh on the hidden gate. In this PR a workaround has been implemented which copies the tensor, using the pool to minimize allocation/de-allocation overhead. An issue has been filed in Make unary ops more efficient with non-contiguous inputs #192 for a more general solution. The issue also affects the sigmoid op, but the overhead is much less.