Rework matrix-related code.
GLSL matrices are stored in column-major order while GL_ARB_vertex/fragment_program
use row-major. So, need to use STATE_MATRIX_TRANSPOSE for built-in matrices.
Unfortunately, this means that the expression M * V isn't very efficient since we
need to extract the rows out of M. And that's the typical expression for vertex
transformation: gl_ModelViewProjectionMatrix * gl_Position.
Solve this inefficiency by looking for M*V expressions and replacing them
with V*Transpose(M).
Also, add support for GLSL 1.20's MatrixTranspose, Inverse and InverseTranspose
matrices.