freedreno/ir3: 16b temps

Notes:

1. first patch will drop out on rebase, that is just so I could do a full CI run
2. Even though this series fixes all the 16b array issues, I should drop the last patch until !5810 (merged) .. switching the pipe-cap causes some regressions in that glsl won't lower large immed arrays to uniforms for us. !5810 (merged) should solve that nicely.
3. We should probably revisit 1b bools in 16b regs, I suspect at least some of the problems that was having should be solved by this series