### What this PR does / why we need it?
Optimize qwen2_vl and qwen2_5_vl.
### Does this PR introduce _any_ user-facing change?
no
### How was this patch tested?
Testing this PR on 1080p picture with tp=1, bs=1 on Qwen2-VL and
Qwen2.5-VL, every fa op's during time lasting from 11ms to 9ms, got
roughly 22% perf boost.
---------
Signed-off-by: zouyida2052 <zouyida@huawei.com>
Signed-off-by: zouyida2052 <zouyida2002@gmail.com>
Co-authored-by: zouyida2052 <zouyida@huawei.com>