Pruning Optimization For Efficient Top-K Document Retrieval With Learned Sparse Representations