Optimize fullword inner loop in nmod_poly_gcd_euclidean

[fredrik-johansson]

Found two more time savers in _nmod_poly_divrem_q1_preinv1_fullword:

• In general r = a + q0*b + q1*c will have high limb larger than n and can even overflow two limbs, but in practice this rarely happens (and almost surely doesn't happen when n is just larger than 2^(FLINT_BITS-1). We can inspect the actual values q0 and q1 before entering the loop to verify that it is safe to do a regular addition instead of a modular addition for the high limb.
• Specialize modular reduction knowing that norm == 0.

The asymptotic improvement is about 15%. This is quite nice since it affects the moduli used by fmpz_poly_gcd_modular.

Effect on nmod_poly_gcd: