PCA recovery thresholds in low-rank matrix inference with sparse noise