better support minibatch sgd about dmlc-core

[tqchen]

I think we did not assume offset[0]==0.

I agree that parser can support IndexType, can you do the change?

[mli]

i can add the indextype.

can you add the interface RowBlock.Slice and RowblockContainer.GetBlock, i can do the implementation and testing. there are two choices, [start_row, end_row), vs (start_row, num_row). either is fine for me.

another issue, is it necessary to have multithread parser in libsvm? i didn't see the speedup from 1 thread to 16 threads in my machine.

[tqchen]

Seems that was due to your improvements in the parser. I do see speedup with 4 threads

[tqchen]

I think we can keep the multi-threading part, and fall back to less thread when necessary

[tqchen]

I have added the RowBlock.Slice, there is no need to add slice to RowBlockContainer, you can do

container.GetBlock().Slice(begin, end)