Suboptimal codegen for quadratic extension

[shamatar]

I was profiling pairing code on different architectures and noticed that this comparison

algebra/ff/src/fields/models/quadratic_extension.rs

Line 278 in 9ce33e6

if P::NONRESIDUE == -P::BaseField::ONE {

is not optimized to nop after monomorphization, most likely because negation is not constant trait. Even though branch predictor should be ok on this place, it's still suboptimal. Most likely the easiest solution until const traits are available is to make additional boolean constant on QuadExtConfig trait that covers such case, or try to provide minus one constant the same way as ONE is provided (and MINUS_ONE is well-defined as there is ONE and Neg + AdditiveGroup) and expect that compiler would optimize-out an equality check (that also is questionable because types are "large")

[shamatar]

I tried a quick local patch to add MINUS_ONE and it eliminates the comparison in the codegen and from profiling. Let me know if you want it to be upstreamed (and it is backward compatible if someone defined e.g. custom MontBackend)