[Regression] Low QWEN3 30B B200 and H100 performance with dropless

[erhoo82]

We have 2X slower performance in QWEN3 30B on H100 and B200 with the dropless schedule compared to the token-drop schedule.

Basically, the perf in 25.11 is 2X slower than 25.09.
We need to fix this in 26.02.

[erhoo82]

@dingqingy-nv to look at only B200 perf

[erhoo82]

@scsudhakaran to help on H100 perf

[erhoo82]

@tomlifu staged this PR for B200 perf. This is mainly about using hybridep.
#2499