Reduce CPU memory usage for `hist` and `approx`.

The approx tree method is rewritten in 1.6 based on the existing `hist` tree method. Although in general, the performance is better, for non-zero gradient it might consume more memory than before (rmse is fine, the new version uses less memory). I have identified a few bottlenecks:

- [ ] Approx only: the transposed `SparsePage` costs a copy of data. This was present in the old version too due to sketching.
- [ ] The gradient index. Right now, on CPU compression only works with dense data, we might add sparse data support.
- [x] The memory buffer for histograms. This may not look like a significant problem, but the builder requires a double buffer and allocates for all nodes. For the bosch dataset, this can cost about 1.5 GB of memory with depth 8 trees.
- [ ] Hist only: we can avoid column matrix is the input if fully dense.