agx: Clean up our uniform register handling

This started with the last patch, keeping load_preamble vectorized to make the NIR less annoying to read. When done naively that introduced some shader-db regressions, which led to the rest of the series to fix properly. The result is a nice win for shader-db overall (especially the reduction in register pressure from the new lowering), a nice simplification to the code, and a nice improvement to the NIR readability.