Switch from statsmodels to scipy #291
Description
We currently only use statsmodels to make KDE's for 1D distributions. From this PBjam constructs an JIT compilable interpolation function representing the KDE (since the statsmodels KDE can't be JIT compiled).
The statsmodels KDE is however very slow to initialize due to the CDF calculation step. This step is significantly faster for the corresponding 1D Gaussian kernel KDE in the Scipy library, which can be adjusted to almost completely mimic the statsmodels implementation in terms of kernel parameters etc.
This wouldn't have any major benefits, but we'd be able to get rid of statsmodels as a dependency which is nice for maintainability and reduce the time to initialize the model classes by a bit.