Extensible call back function for DMatrix.

[trivialfis]

This is an issue for tracking initializing data on distributed environment. On both Spark side and Dask side we need to concatenate partitions into a single block of data to initialize the DMatrix. Even with the new DeviceQuantileDMatrix the peak memory usage is still twice the size of actual data.

One solution is to implement some sorts of call back functions for internal data adapter to process these partitions inplace to avoid concatenating data.