[Speed] speed up the sequence_expand op

[dzhwinter]

The sequence_expand op contains multiple cuda kernel function calls, that means for every pair in lod scope, it needs to launch the cuda kernel for one time. In fact, they can merge to one, which will speed up the sequence processing hugely.