Optimize row reduction with delayed canonicalisation in more cases

[fredrik-johansson]

• fmpz_mod, fq, fq_nmod and nf should implement lu_classical_delayed with delayed canonicalisation (see nmod, mpn_mod)

• mpn_mod and fmpz_mod should overload gr_mat_reduce_rows with delayed reduction (similar to nmod_mat_reduce_rows). Note that there is an fmpz_mod_mat_reduce_rows which currently doesn't do delayed reduction. Note: could do this for nf also.

• Check if nmod_mat_minpoly is fast enough when based on generics if gr_mat_reduce_rows wraps nmod_mat_reduce_row for nmods, so that we can make nmod_mat_minpoly a wrapper around gr_mat_minpoly_field

• Independent of the above, gr_mat_reduce_rows_generic should use vec operations for the inner loops