The implementation uses a trait to fold over tuples, summing the
dimensions in one direction and checking for equality in the other, and
then uses `fixed_{rows,columns}_mut` if the dimensions are static, or
`{rows,columns}_mut` if the dimensions are dynamic, together with
`copy_from` to construct the output matrix.

…onst parameters to the end of declarations to support the rustc version used by the cuda test suite.

…HStack`/`VStack` to facilitate implementing lazy stacking.

…rmediate allocations when building a matrix with a mix of horizontal and vertical stacking.