bugfix: Fix the behavior of decode cuda graph wrapper

[yzh119]

In this PR, we fixed the grid size of the kernels launched by a decode cuda graph wrapper:

• We always pad the block size to a fixed value.
• Early stop for threadblocks that is out of bound.