------------------
- Add NumPy as a mandatory dependency.
- Only compute matrices in C code and move sorting the values, casting to double, and computing the actual breaks to Python/Cython code for better maintainability.
- Improve performance by using 1D arrays instead of 2D arrays in ``JenksBreakValues`` C function.
- Preserve the precision of the original list/array of values in the returned breaks.
- Fix bug when requesting a number of class equal to the number of values.
- Raise an exception when the number of classes is greater than the number of unique values (however this might change in the future by choosing to return a list of breaks shorter than the one requested by the user).
- Rename ``nb_class`` parameter to ``n_classes`` (notably to be closer to sklearn ``n_clusters`` parameter).