Vision Attention

[Fxycst1213]

#4509 When implementing InternVL-VIT, we found that using Vision Attention requires modifying Vision Attention.py, which makes the code less elegant. Since InternVL-VIT uses qk_norm, we decided not to use Vision Attention. Should we modify the implementation of qk_norm in Vision Attention?