Output precision lowering

Landing more of the mediump work. This enables support for fp16 in the GLSL IR and lowers output variables to fp16. On the IR3 it adds support for folding conversions into producing instructions; enough that we can typically generate fp16 outputs without adding instructions.

This also lands the conversion opcodes that we've agreed on, in particular the f2fmp conversion that solves some of the issues about when we can eliminate conversions