nir: add unordered comparisons, ordered fne, ford and funord, remove aco's optimizations for them

AMD and Nvidia both have matching hardware for all of these opcodes. Until now, aco had a few patterns for this in aco_optimizer, but with NIR opcodes the optimizations can be expressed in just a few nir_opt_algebraic patterns. Having a NIR option is also helpful for @idr's MR that optimizes some of these patterns for hardware that does not support ordered fne.

This MR also contains two nir_opt_algebraic changes that were needed to remove the backend optimizations without regressions.