gallivm: Fix up constant-index nir array register accesses.
Follow-up to !21001 (merged) that actually splits out cleanly, getting those tests to be nice and quick once both are merged:
Instead of generating num_components*simdwidth scattered stores, if
there's no indirect then we can just look up the pointer to the
base_offset and do a simd store there.
dEQP-VK.subgroups.ballot_broadcast.compute.subgroupbroadcast_i64vec4 goes
from 30s to ~2s.