[SPARK-5016] [MLLIB] Distribute GMM mixture components to executors

Distribute expensive portions of computation for Gaussian mixture components (in particular, pre-computation of `MultivariateGaussian.rootSigmaInv`, the inverse covariance matrix and covariance determinant) across executors. Repost of PR#4654.

Notes for reviewers:
 * What should be the policy for when to distribute computation. Always? When numClusters > threshold? User-specified param?

TODO:
 * Performance testing and comparison for large number of clusters

Author: Feynman Liang <fliang@databricks.com>

Closes #7166 from feynmanliang/GMM_parallel_mixtures and squashes the following commits:

4f351fa [Feynman Liang] Update heuristic and scaladoc
5ea947e [Feynman Liang] Fix parallelization logic
00eb7db [Feynman Liang] Add helper method for GMM's M step, remove distributeGaussians flag
e7c8127 [Feynman Liang] Add distributeGaussians flag and tests
1da3c7f [Feynman Liang] Distribute mixtures