Make *_poly_mulhigh available and as efficient as *_poly_mullow

[fredrik-johansson]

Mulhigh for polynomials is exactly the same operation as mullow, just reindexed, so we ought to make it exactly as performant in all cases.

Upside: in Newton division and reduction with precomputed inverses, we could just do a mulhigh instead of reversing the operands, doing a mullow, and reversing the result.

[albinahlback]

+1, in case this was not indicated enough by my reaction already