ENH: rewrite of stats.ppcc_plot

[rgommers]

Closes Statistics Review issue gh-679.

Also some minor fixes to stats.probplot and stats.tukeylambda are included.

[rgommers]

This PR should be pretty straightforward, any takers for a review?

[WarrenWeckesser]

I haven't reviewed the ppcc_plot changes, but I noticed that the tukeylambda PPF function could be simplified to boxcox(q, lam) - boxcox1p(-q, lam), where boxcox and boxcox1p are from scipy.special. If you'd like, you can remove your change to tukeylambda._ppf, and I'll create a PR to make this change.

[josef-pkt]

Reading through the code: The changes look good to me, I don't see any problems and it improves the code.

I haven't paid any attention to what the empty array behavior is now in various functions.
I guess if there is only one point, or maybe a few, the plot and fit will break down. Is probplot defined for very small arrays?

[rgommers]

probplot seems to work for any length input:

• length 0 is special cased
• length 1 gives a RuntimeWarning and returns ((array([ 0.]), array([1])), (nan, nan, 0.0)).
• length 2 and up return values and don't give any warning. There's not much point of course for a length-2 input, but it seems overkill to add warnings for specific input sized there (also not clear where you'd put the cutoff then).

Plots are all sensible as well, except a plot for length-1 which of course is just a single point in a figure.

For empty input ppcc_plot simple inherits the probplot behavior. At some point we may want to go through and review empty input handling for consistency between all stats functions, but that's not something I want to do now or in the near future.

[rgommers]

@WarrenWeckesser that sounds good. Although easier, why not merge this PR now (since Josef just reviewed it) and then you make your proposed change on top of it?

[WarrenWeckesser]

@rgommers: That's fine, too--less work for you, which is good. :)

[rgommers]

Great, thanks. Merging.

[WarrenWeckesser]

tukeylambda ppf change is here: #4692