Reject low-order points by checking projective coordinates directly
instead of using affine coordinates.

Equivalent, but saves CPU cycles (~254 field multiplications total
before, 3 field multiplications after).