Tight Data-Robust Bounds to Mutual Information Combining Shuffling and Model Selection Techniques