Use delayed canonicalization in fmpz_mod_mat_reduce_row

[fredrik-johansson]

Sample timings for minpoly of an n x n randtest matrix over Z/pZ with a 400-bit p:

n       OLD         NEW
2       1.08e-05    1.06e-05
4       1.67e-05    1.59e-05
8       6.12e-05    4.76e-05
16      0.000403    0.000233
32      0.003       0.00139
64      0.0241      0.0104
128     0.197       0.084
256     1.68        0.799
512     12.573      5.693

One sub-task of #2178.

[fredrik-johansson]

Update: also implemented for mpn_mod.

Timings for mpn_mod matrix minpoly with 400-bit modulus:

n       OLD         NEW
2       3.49e-06    2.43e-06
4       1.11e-05    9.16e-06
8       4.99e-05    4.03e-05
16      0.000301    0.000204
32      0.00215     0.00129
64      0.0163      0.00894
128     0.128       0.0667
256     1.006       0.529
512     8.074       4.1

100-bit modulus:

n       OLD         NEW
2       1.4e-06     1.6e-06
4       6.01e-06    7.12e-06
8       1.03e-05    1.11e-05
16      4.96e-05    4.2e-05
32      0.000457    0.00021
64      0.00356     0.00122
128     0.0281      0.00821
256     0.223       0.0589
512     1.751       0.45

[albinahlback]

That's a big speedup!

[fredrik-johansson]

I wish I knew why the 32-bit CI is now failing in the tests for gr_poly (presumably just because of the change to gr_ctx_init_random generating different random inputs?). I can't reproduce that error on a different 32-bit system.

[fredrik-johansson]

The segfault is as far as I can tell unrelated to the contents of this PR other than the change in randomization, so I will open a separate issue.