Ndd

Added
- `MillerMadow` estimator class
- `AsymptoticNSB` estimator class
- `Grassberger` estimator class
Changed
The signature of the *entropy* function has been changed to allow
arbitrary entropy estimators. The new signature is

entropy(pk, k=None, estimator='NSB', return_std=False)

Check `ndd.entropy_estimators` for the available estimators.

Changed
For methods/functions working on data matrices:
the default input is a **n-by-p** 2D array (n samples from p discrete
variables, with different samples on different **rows**).
Since release 1.3, the default was a transposed (**p-by-n**) data matrix.
The behavior of functions taking frequency counts as input
(e.g. the `ndd.entropy` function) is unchanged.
Added
- builds on Windows (with MinGW-W64)
- builds on MacOS (thanks to https://github.com/ccattuto)

Added
- `ndd.kullback_leibler_divergence`

Changed
- `r` (length of combinations) defaults to None

Changed
- input data arrays must be p-by-n 2D ndarrays containing
n samples from p discrete variables. This affects all methods/functions
working directly on data:
- histogram
- from_data
- interaction_information
- coinformation
- mutual_information
- conditional_entropy

Changed
- fixed conditional_entropy function
- histogram: `axis` should be None is data matrix is tarnsposed.