v3d: backend VS store lowering in NIR (needs nir review)

This moves most of the backend logic for how VS stores work into NIR, so that NIR scheduling can see it and nir_intrinsic_store_output can emit the store directly to reduce register pressure (see the nir-schedule branch for why I'm trying to do this)

For NIR reviewers, I'm looking for review on patch 1, and an ack on the new intrinsics and the general strategy of patch 3.