turn the cryptically named scipy.special information theory functions into ufuncs

[argriffing]

These ufuncs try to be careful about corner cases, and none of them can be directly replaced by existing ufuncs functions like xlogy. The entr(x) ufunc is very similar to -xlogy(x,x), but it has different behavior at negative x. The other two ufuncs are convex on all (x, y) whereas their earlier implementations that were clever enough to handle the limit at the origin still returned nan for some x,y points.

These functions are intended to have the same definitions as the functions of the same name in "disciplined convex programming" (http://stanford.edu/~boyd/papers/disc_cvx_prog.html) projects like cvx and cvxpy (note that these are different from cvxopt). The hope is that because these functions are so carefully defined, they can simplify some of the corner case headaches in higher level scipy functions related to statistics and divergences.

[argriffing]

Running generate_ufuncs.py seems to have caused some changes related to earlier PRs. Some typos are fixed and some "versionadded should be in the Notes section" changes are made.

[coveralls]

Coverage decreased (-0.01%) when pulling fbc76f5 on argriffing:infotheory-ufuncs into 8b8531d on scipy:master.

[coveralls]

Coverage decreased (-0.01%) when pulling fbc76f5 on argriffing:infotheory-ufuncs into 8b8531d on scipy:master.

[pv]

Maybe stuff all of them into a single .pxd file? There's no technical
need to have one .pxd per one function.

[argriffing]

@pv OK I put them in to a _convex_analysis.pxd file to emphasize their focus and explain the choices for corner cases or weird parts of the domain. For example, these functions are defined in a context that may depend on the concavity of log(x) more than on its characterization as the inverse function of \sum_{k=0}^\infty{x^k / k!}, so the functions don't deal with imaginary numbers.

[argriffing]

test run exceeds 50 minutes

[argriffing]

Added the Huber loss function as a ufunc. This element-wise convex function is already used in scikit-learn, cvxpy, and statsmodels.

[coveralls]

Coverage decreased (-0.01%) when pulling 33ff802 on argriffing:infotheory-ufuncs into 8b8531d on scipy:master.

[argriffing]

Added the convex Pseudo-Huber function as a ufunc. This is used in statsmodels and could be used in scikit-learn. It's a smoother variant of the Huber function.

[coveralls]

Coverage decreased (-0.0%) when pulling be7f6b2 on argriffing:infotheory-ufuncs into 8b8531d on scipy:master.

[ev-br]

LGTM, merging.