Is add_no_canonicalize_trashing_input needed for goldilocks reduce128?

[chriseth]

In the implementation of reduce128 for goldilocks, we use add_no_canonicalize_trashing_input and after that we call wrap' again. The code was taken from an implementation where it does not matter that the result is reduced modulo the field prime, as long as it fits 64 bits. We generally assume that field elements are always stored in canonical form (i.e. less than the field prime). Because of that I introduced the wrapcall at the end. So the question is if theadd_no_canonicalize_trashing_input` is that useful or if it can be implemented in a better way.

Note that we now have two files with similar impls for goldilocks.