Implementation tasks/milestones and possible solutions

[siftheads1]

1. GetGradient: How to do it for subsamples
   ->Compute entire gradient and zero-out v.s. only compute for sampled data(we believe the latter is a better approach)

2. GHistIndexMatrix(part where maps 'gradient values' to bin index) -> How to do it for subsamples

3. Building Histogram -> How to update histogram with newly sampled data instead of building from scratch
   (Possible solution: modify logic that builds "row_indices"(target row index of histogram building logic(RowsWiseBuildHistKernel)

4. Find slicing method for DMatrix(probably needed to complete the objectives above)
   -> We assume this can be done by appropriately slice MetaInfo and GHistIndexMatrix.
   Need to find methods to do the tasks above.