Add Mixture Density Network as built-in CondDensityEstimator implementation

[iwahle]

Conditional density estimation can be done by a variety of learning models. While the current software includes an implementation for estimating the expectation of the conditional density using a neural network, a mixture density network (Bishop 1994: https://publications.aston.ac.uk/id/eprint/373/1/NCRG_94_004.pdf) could be used to learn a more precise estimate of the conditional distribution with sufficient data.
This feature addition would involve implementing a new class that inherits cfl.cond_density_estimation.cde_model.CDEModel so that it matches the interface that CondDensityEstimator expects.