ac/nir: Implement packed 16-bit outputs for last pre-rasterization stages.

Implement packed 16-bit outputs for VS/TES/GS when they are the last pre-rasterization stage.

Note that these stages only supported packed 16-bit outputs in the OpenGL specific slots that were dedicated to mediump. To keep the code simple, I decided to handle this case separately, without disturbing the parts that handle the mediump slots.