Understanding K-Means Clustering on Matrix Data: A New Approach for High-Dimensional Observations