Vector rational reconstruction

[fredrik-johansson]

There should be functions to reconstruct an _fmpq_vec and fmpq_poly from a modular image, like fmpq_mat_set_fmpz_mat_mod_fmpz.

Note that one can use different strategies. Calling fmpq_reconstruct_fmpz in a loop is good if the coefficients have independent denominators. Trying to track the common denominator as in fmpq_mat_set_fmpz_mat_mod_fmpz is efficient if one has a small common denominator. We could try to figure out (and carefully profile) some heuristic to switch between these strategies.

Another thing worth optimizing is the numerator/denominator bound which by default is taken to be balanced. If we detect small denominators heuristically and allow using most of the bits for the numerator, we can potentially get a 2x speedup for operations like Dixon solving and polynomial interpolation in the common case where the (common) denominator is much smaller than the numerators.