Created by: Superjomn

Current LoD stores end offsets of each sentence, that design might save space (a little).

In dynamic-batch RNN (that take variant length sentences as input) need to sort the elements in one level according to number of elements in lower level, in that case, the LoD may be reordered, that need a begin offset in addition to end offset.

So a pair of begin and end offsets should be introduced to LoD's implementation.