Vectorization of Anderson-Darling computations.

[anntzer]

Results in >1000x acceleration of scipy.stats.anderson_ksamp for
datasets of N~1000.

[rgommers]

that's a nice speedup:)

[rgommers]

Would be good to add a benchmark for this function, so we'll detect any regressions. Should be easy to do - example: gh-4621

[josef-pkt]

just out of curiosity: What's the usecase with N=1000?
I always thought of k-sample tests as test with just a few, like 3 to 10, samples.

[anntzer]

Actually I just wanted to run a 2-sample test, but with a large number of data points in each sample. Moreover, the differences between the distributions would be mostly in the tail region, suggesting the use of something else than the KS test.

[josef-pkt]

I mixed up k and N in my reading, so large N (total nobs) is not a surprise.

[rgommers]

Thanks @anntzer, merged.