The comparer function for COMPLEX types wasn't properly optimized
during cross-compilation because the types of REALPART and IMAGPART
weren't derived by the cross-compiler on account of
CROSS-FLOAT-INFINITY-KLUDGE, but the derivers do not actually perform
any mathematical derivation, just
(complex (double-float 10d0)) => (double-float 10d0).
Enabling them during cross-compilation allows = on complex floats be
optimized and avoids consing.