Use better initial guesses for Roots

[cuviper]

When a BigUint is small enough to fit in a u64, we use that to compute roots instead.

Otherwise, we can get a closer approximation of roots via f64. For values that are even too big for that, we can shift them down in multiples of n enough that they do fit, and then shift the result back and refine the answer further.

Benchmarked with an i7-7700K on Fedora 29, rustc 1.32.0-nightly (21f268495 2018-12-02).

 name             roots-master ns/iter  roots-opt ns/iter  diff ns/iter   diff %   speedup
 big1k_cbrt       18,787                6,934                   -11,853  -63.09%    x 2.71
 big1k_nth_10     20,896                4,595                   -16,301  -78.01%    x 4.55
 big1k_nth_100    53,912                1,326                   -52,586  -97.54%   x 40.66
 big1k_nth_1000   4,175                 5,205                     1,030   24.67%    x 0.80
 big1k_nth_10000  7,634                 23                       -7,611  -99.70%  x 331.91
 big1k_sqrt       19,083                9,804                    -9,279  -48.62%    x 1.95
 big2k_cbrt       46,799                20,917                  -25,882  -55.30%    x 2.24
 big2k_nth_10     27,959                12,296                  -15,663  -56.02%    x 2.27
 big2k_nth_100    77,422                7,430                   -69,992  -90.40%   x 10.42
 big2k_nth_1000   10,851                10,611                     -240   -2.21%    x 1.02
 big2k_nth_10000  7,963                 24                       -7,939  -99.70%  x 331.79
 big2k_sqrt       53,835                20,542                  -33,293  -61.84%    x 2.62
 big4k_cbrt       117,757               45,320                  -72,437  -61.51%    x 2.60
 big4k_nth_10     87,084                32,310                  -54,774  -62.90%    x 2.70
 big4k_nth_100    43,477                22,657                  -20,820  -47.89%    x 1.92
 big4k_nth_1000   101,478               102,153                     675    0.67%    x 0.99
 big4k_nth_10000  8,384                 24                       -8,360  -99.71%  x 349.33
 big4k_sqrt       142,891               48,711                  -94,180  -65.91%    x 2.93
 big64_cbrt       2,417                 59                       -2,358  -97.56%   x 40.97
 big64_nth_10     4,595                 84                       -4,511  -98.17%   x 54.70
 big64_sqrt       1,109                 37                       -1,072  -96.66%   x 29.97

Everything is a big win except nth_1000, but even the loss there is fast enough that I'm not worried. I doubt nth_root(1000) would realistically be used anyway -- I just wanted challenging cases.

[cuviper]

bors r+

[bors]

Build succeeded

• continuous-integration/travis-ci/push