[R] Move gc data protection to R side

[david-cortes]

ref #9810
ref #11092 (comment)

This PR moves the logic for protection of R data from the garbage collector towards the R side.

I'm not sure whether data holding fields of the DMatrix also needs to be kept protected after setting it in the proxy dmatrix, but this PR does so just in case.

[trivialfis]

Let me try to create a prototype for the #11088 and see how this should work. I need to store the categories somewhere.

[trivialfis]

Should we revert #11092 before applying this PR? The previous one seems redundant now.

[david-cortes]

Should we revert #11092 before applying this PR? The previous one seems redundant now.

It is undoing the changes from it already.

But it leaves me wondering: what's the difference between xgb.ExtMemDMatrix and xgb.QuantileDMatrix.from_iterator? Both of them seem to apply exclusively to histogram-based models, and both of them seem to access the data the same amount of times.

Is xgb.ExtMemDMatrix perhaps meant to leave the iterator alive throughout the call to xgb.train?

[trivialfis]

The xgb.ExtMemDMatrix stores a cache on disk, whereas the QDM.from_iterator concatenates everything in memory and saves memory by using the compressed histogram bin index instead of raw input.

In addition, the former applies to both approx and hist, while the latter is hist only.

https://github.com/dmlc/xgboost/blob/master/doc/tutorials/external_memory.rst#compared-to-the-quantiledmatrix

[trivialfis]

I'm not sure whether data holding fields of the DMatrix

If you meant things like labels and weights, no, they don't need any protection.

[david-cortes]

I'm not sure whether data holding fields of the DMatrix

If you meant things like labels and weights, no, they don't need any protection.

Thanks, modified the PR to not keep those protected after setting them.

[trivialfis]

Will take a deeper look tomorrow.