Using Subset Log-Likelihoods to Trim Outliers in Gaussian Mixture Models