When constructing a matrix from another matrix with smaller dimensions,
there's no need to extract the scalars out of columns and rebuild the
resulting matrix from scalars - instead, we can just construct shorter
vectors with OpShuffle and combine them to the final result.

This keeps the common casts such as mat3(mat4) in vector registers,
which may improve performance for some GPUs, and cleans up output of
translation tools like SPIRV-Cross.

Fixes #1412.