Multi-Source Knowledge Pruning for Retrieval-Augmented Generation: A Benchmark and Empirical Study