Kde boundaries

[aloctavodia]

This KDE implementation is faster than scipy.gaussian_kde. According to my test between ~10 times faster (for ~1000 points) and ~40 times faster (for ~10000 points). This implementation includes a boundary correction (using the reflection method). I think the only caveat is that it will not work very nice if the number of data points is low (probably lower than ~1000/500), in such a cases the tails of the distributions are not estimated very nice. Nevertheless I think this is OK, given that in general people uses at least 1000 samples.

The next image is an example of a KDE estimation without boundary corrections (the SciPy implementation)

And this is an example with boundary correction (the implementation in this PR)

I know is a small detail, but in my experience KDE plots without the correction can be confusing to those unaware of how KDE estimation works, or at least those unfamiliar with PyMC3/scipy-KDE.

This is an adaptation of the code from this repository and made some modifications. I made this change following this suggestion arviz-devs/arviz#5

[springcoil]

Hey thanks for this commit. I agree with your comments about KDE.

This looks a good contribution.

[twiecki]

Nice, that looks way better!