nir: Optimize nir_validate

Profiling the CTS, nir_validate is surprisingly hot -- about 10% of the time on my M1 GLES job. This contributes to slow CI, so let's optimize the 💩 out of NIR validation to speed up testing.

The goal of this series is to leave the core algorithms for validation unchanged, without dropping coverage, because validation in CI is a Good Thing. Instead we seek improvements by implementing the existing code more efficiently. Details per-patch, there's no overall structure here, more of a potpourrie of perf positivity.

Overall, with a properly configured shader cache cleaned in between runs (to emulate the CI setup), deqp-runner reported time for the M1 GLES job decreases from 2:52 to 2:41 across the series. 6% reduction in CI testing time for one small MR is nothing to sneeze at 🤧