Learning from Label Proportions: A Mutual Contamination Framework