vkd3d-shader/hlsl: Lower matrix swizzles.

As @nsivov pointed out, we are currently parsing matrix swizzles (e.g. mat._m12_m32) and creating an hlsl_ir_swizzle for the matrix itself. We are currently allowing shaders that use these to compile even though they are not working.

While our compilation passes don't expect hlsl_ir_swizzles with matrix values, turns out we need them as a parse-time construct because we don't know if they will be used as the lhs of an assignment during parsing, similarly to hlsl_ir_indexes.

This patch adds tests and a pass to lower these matrix swizzles.

Implementing assignments with matrix swizzles is still pending.