freedreno: UBO pointers fix (and NIR num_ubos fix)

Rob wanted to fix UBOs before my LDC changes landed, but even after his series we have some broken pointer math in ir3_compiler_nir on older chips, and sensible asserts in fixing that broken pointer math showed that we were doing some weird stuff that seemed to only work out by chance (indirecting on register number minus 2!)

Take a step back and fix the NIR that our lowering pass produces so that it's consistent. In doing so, I found that the NIR we were being passed in was inconsistent in one case, so fix that too!