For the classical GMM trainer,
Improved the interface and overall computing accuracy and efficiency for extremely high dimensional problems.
Added the option of embedding small positive constants on the diagonal entries of covariance matrices.
Added the option of running K-means and K-means++ beforehand.
Adjusted the method of copying bits to avoid warning messages from 32-bit compilers.
Enabled classical GMM trainer to fix any of the model parameters, including mixture weights, means and covariances of the Gaussian components, during training.