Generic mpn_mullow and hardcoded ones for x86

[albinahlback]

Edit: Also solves #1879

[fredrik-johansson]

BTW, it might be possible to avoid the need for a carry-out. I will make an issue about that.

[albinahlback]

Hmm, I think we should investigate this before we go further. While hardcoded routines are sort of easy for me, the generic ones are considerably harder for me. Would be nice to just utilize GMP's mpn_mullo_basecase if it is possible.

[albinahlback]

@fredrik-johansson

        mullo_basecase / FLINT || mul / mullow
n =  1:            1.20x       ||     1.00
n =  2:            1.16x       ||     1.14
n =  3:            0.92x       ||     1.32
n =  4:            1.16x       ||     1.28
n =  5:            1.26x       ||     1.27
n =  6:            1.30x       ||     1.30
n =  7:            1.22x       ||     1.29
n =  8:            1.23x       ||     1.32
n =  9:            1.16x       ||     1.42
n = 10:            1.13x       ||     1.47
n = 11:            1.08x       ||     1.49
n = 12:            1.06x       ||     1.50
n = 13:            1.05x       ||     1.50
n = 14:            1.03x       ||     1.54
n = 15:            1.03x       ||     1.37
n = 16:            1.02x       ||     1.40
n = 17:            1.00x       ||     1.49
n = 18:            1.00x       ||     1.45
n = 19:            0.99x       ||     1.53
n = 20:            1.00x       ||     1.45
n = 21:            0.99x       ||     1.54
n = 22:            1.00x       ||     1.47
n = 23:            0.99x       ||     1.53
n = 24:            0.99x       ||     1.47
n = 25:            0.99x       ||     1.51
n = 26:            0.99x       ||     1.44
n = 27:            0.99x       ||     1.50
n = 28:            0.99x       ||     1.46
n = 29:            0.99x       ||     1.40
n = 30:            0.99x       ||     1.30

I.e. very good on Skylake when taking into account that we calculate an extra limb.

[albinahlback]

Valgrind looks OK for this, so I'll merge!