Use MatrixTransposeMatrix multiply from small_blas.h

Prior to this Cheng Wang had to work the MatrixTransposeMatrixMultiply
function in inner_product_computer because of its lagging performance.

Now that the implementation of MatrixTransposeMatrixMultiply
has been updated to match the implementation in inner_product_computer
this CL removes the redundancy.

This CL also includes a fix to MatrixTransposeMatrixMultiply
which was missed in the earlier refactoring, because the test
coverage for small_blas is broken. The tests for InnerProductComputer
caught the problem.

Change-Id: Ia8890010da058c2d9fe738dc77f0be34af5618ab