scikit-learn: Improve GMM module

by Wei Xue for Python Software Foundation

Current implementation of Gaussian mixture model (GMM), variational Bayesian Gaussian mixture model (VBGMM) and Dirichlet process Gaussian mixture model (DPGMM) in scikit-learn have many problems. They suffer from interface incompatibility, incorrect model training and incompleteness of testing. In this project, I propose to clean up APIs of GMM, re-implement the updating functions of VBGMM and DPGMM based on well-acknowledged literature, and pass various test cases.