SRFormerV2: Taking a Closer Look at Permuted Self-Attention for Image Super-Resolution