nir: New nir_load_ubo_vec4 intrinsic (major freedreno codegen win)

Here, I extract some ir3 compiler backend code to do a new core intrinsic for vec4-unit UBO loading, document some common restrictions on UBO loading, and use those restrictions to get us a massive win in freedreno codegen. This was also needed in !3395 (merged) to fix one of the few remaining regressions.

This may also be of interest to @gerddie as there's some r600 backend lowering with a comment ("encode the start component") that sounds like it wants this lowering instead.