Revert gh-5938, restore ks_2samp

[ev-br]

As discussed in gh-6435,

1. Cythonize stats.ks_2samp for a ~33% gain in speed. #5938 changed ks_2samp for scipy 0.18.0,
2. the change was supposed to be a moderate speed up only, but it changed the behavior in case of ties in the data
3. both "old" (0.17.0) and "new" (0.18.0) behavior is somewhat ad hoc, and the "new" is not clearly better than the "old"
4. figuring out what is the obviously correct behavior is not trivial and nobody seems to have time/energy for it.

Therefore, this PR simply reverts gh-5938 and restores the implementation of stats.ks_2samp from scipy 0.17.0 so that we keep backwards compatibility until someone steps up for 4. above.

Points to consider:

• scipy.stats.ks_2samp returns different values on different computers #6435 has an example which can be used for testing the result against scipy 0.17.0 and older. ISTM a test like this is going to be confusing in the long run, but I'd be fine adding one if there are strong opinions.
• So far I kept the cython implementation of ks_2samp in _stats.pyx which is there but unused. Do we want to keep it in the source code, or we're fine with it surviving in the git history only?

[rgommers]

ISTM a test like this is going to be confusing in the long run, but I'd be fine adding one if there are strong opinions.

Agreed, if it's not clearly correct, then it probably doesn't have too much value.

[rgommers]

So far I kept the cython implementation of ks_2samp in _stats.pyx which is there but unused. Do we want to keep it in the source code, or we're fine with it surviving in the git history only?

I suggest to remove it and open a new issue that explains the history and what has to be done to get the faster version of ks_2samp back. Keeping dead code around is just going to waste someone's time later on, figuring out why it's there.

[ev-br]

OK, dead code is gone.

[rgommers]

In it goes then. Thanks @ev-br

[ev-br]

backported in gh-6557