Add modular splitting

[David-Berghaus]

Added modular splitting as described in Brent, Zimmermann: Modern Computer Arithmetic, Section 4.4.3.

Remarks:
-I ended up not using set_shallow because the swap-approach has the convenience that all coefficients that are out of bounds are simply zero entries which are ignored by the dot product routines.
-I started working on gr_poly_evaluate_other_modular but stopped since I noticed that gr_dot_other still uses the naive method. We therefore cannot get a performance advantage over gr_poly_evaluate_other_rectangular but potentially worse numerical stability.

[fredrik-johansson]

This will be a nice algorithm to have!

The code needs a little bit of work. You can't safely swap entries from poly because this is a read-only argument: it may be read simultaneously from another thread, for example. So tmp should actually be allocated shallowly and set_shallow should be used. See gr_mat/nonsingular_solve_tril.c for an example. For the dot product where you're currently zero-padding, you should be able to calculate the actual length.

Indeed, this algorithm could be implemented more cleanly if gr_dot supported strided access which is unfortunately not the case.

-I started working on gr_poly_evaluate_other_modular but stopped since I noticed that gr_dot_other still uses the naive method. We therefore cannot get a performance advantage over gr_poly_evaluate_other_rectangular but potentially worse numerical stability.

Yes, this will make sense to add when we have an overloadable gr_dot_other / gr_other_dot.

[David-Berghaus]

Oh yes, I was stupid and didn't think of potential multi-threading!

I replaced swap with set_shallow now.

[fredrik-johansson]

Just some minor style issues to fix: use /* */ comments, { on new lines.

[fredrik-johansson]

BTW, gr_fmpz_poly_evaluate_modular is documented but does not seem to be part of the PR.

[David-Berghaus]

Done and fixed!

[fredrik-johansson]

Nice!

[fredrik-johansson]

I pushed a small optimization: the variable tmp_gr is not needed if you compute one more entry in the x vector.