Increase correctness testing for BATCH_MATMUL dimensionality reduction

[ddavis-2015]

@tensorflow/micro

The existing BATCH_MATMUL dimensionality reduction code does not fully test the correctness of that code. Specifically there are no tests for a LHS tensor with shape [..., N, A, B] where N > 1, and a RHS tensor with shape [..., 1, B, C]. Current tests only check for N = 1. This means that any change to the code involving the row/column transpose of these tensors may pass successfully, even though results for N > 1 would be incorrect.

Existing rank 3 tensor unit tests for BATCH_MATMUL should have the LHS tensor shape batch size be N > 1 for each of the four possible adjoint parameter settings.