Machine-dependent use of n_precomp functions

[fredrik-johansson]

On my machine n_*_preinv functions are often faster than the n_*_precomp versions, e.g. a 54-bit primality test is faster than a 52-bit one because of this.

We ought to profile these on various current machines and choose the precomp functions conditionally based a flag that we define in flint-mparam.h.

[fredrik-johansson]

One could also see if the precomp versions run faster when reimplemented using floating-point FMA. However, keeping residues in double format through a whole algorithm as in fft_small ought to be better than converting back and forth between ulong and double.