Use pykdtree when available

[QuLogic]

Rationale

The pykdtree project by @storpipfugl is super fast compared to scipy, so this tries to use it if it's available.

Implications

There is no effect on tests, but a great effect on the examples:

                   example   function     master   pykdtree     ratio
0          aurora_forecast    imshow0   1.614568   0.697115  0.431765
1          aurora_forecast    imshow1   1.620853   0.704454  0.434619
2          aurora_forecast       draw   0.916317   1.119499  1.221738
3        eccentric_ellipse     imshow   2.400456   1.269291  0.528771
4        eccentric_ellipse       draw   0.341589   0.677258  1.982666
5   effects_of_the_ellipse  add_image   0.000008   0.000008  1.031250
6   effects_of_the_ellipse       draw   1.835607   1.980497  1.078933
7             eyja_volcano  add_image   0.000003   0.000003  0.928571
8             eyja_volcano       draw   3.526339  10.100620  2.864336
9            geostationary     imshow  96.402943   1.546087  0.016038
10           geostationary       draw   0.643935   0.896134  1.391653
11                     wms    add_wms   0.514387   0.521981  1.014763
12                     wms       draw   2.045766   2.763029  1.350609

In almost all cases, the imshow is approximately half the time while draw is slightly slower. For some reason, the draw in eyja_volcano is much slower which I need to investigate a bit, but the real win here is geostationary which is two orders of magnitude faster.

This speedup should hopefully improve doc builds a bit (at least where it was added.)

[QuLogic]

Oh, actually, I think that eyja_volcano one is a fluke. I re-ran everything 5 times and took averages:

                                  master_mean  master_std  pykdtree_mean  pykdtree_std    ratio  ratio_err
example                function                                                                           
aurora_forecast        draw          1.678460    0.115361       1.657326      0.326995 0.987408   0.206300
                       imshow0       1.620274    0.034479       0.701028      0.009098 0.432660   0.010784
                       imshow1       1.596412    0.003287       0.717052      0.014395 0.449165   0.009064
eccentric_ellipse      draw          1.033083    0.433136       0.838907      0.155962 0.812043   0.372432
                       imshow        2.452178    0.067519       1.273574      0.009867 0.519364   0.014856
effects_of_the_ellipse add_image     0.000008    0.000001       0.000008      0.000001 0.988636   0.160593
                       draw          2.272155    0.206711       2.140229      0.113149 0.941938   0.099112
eyja_volcano           add_image     0.000003    0.000000       0.000003      0.000000 0.955224   0.071966
                       draw          8.419627    9.380086       4.448825      0.128522 0.528387   0.588861
geostationary          draw          0.889276    0.086658       0.927128      0.150869 1.042565   0.197748
                       imshow       98.028681    1.924029       1.555004      0.018169 0.015863   0.000362
wms                    add_wms       0.508353    0.039948       0.501570      0.018772 0.986658   0.085879
                       draw          2.873218    0.163422       2.394889      0.180951 0.833521   0.078828

Most imshow are ~50%, draw is between ~5% under and 5% over (but it's generally small and lots of variance anyway), and geostationary is still the most amazing running with just 1.5% of the original time.

Docs are 2.5 minutes faster, so maybe we should add pykdtree on both builds?

[QuLogic]

This change is there because we don't really care about the distance and pykdtree returns differing results for out-of-domain vs. nan storpipfugl/pykdtree#35

[QuLogic]

Rebased; I enabled pykdtree on both doc builds because a) it's so much faster, and b) there's already checks for plain scipy on Travis.

[ajdawson]

I don't see any reason why we wouldn't want to do this, so 👍