Use RMM's pached CCCL

[jakirkham]

Make sure to search for RMM if it will be used. This should pick up the patched CCCL from RMM.

If RMM is not being used and this is a CUDA build, search for CCCL explicitly.

[jakirkham]

cc @trivialfis @bdice

[jakirkham]

Thanks Hyunsu! 🙏

Applied a similar change to 2.1.x with PR: #11353

[jakirkham]

Tested a backport of this change to XGBoost 2.1.x

Tried this in RAPIDS and confirmed RMM is now picked up: rapidsai/xgboost-feedstock#85 (comment)

Also confirmed building without RMM (like in conda-forge) works unchanged: conda-forge/xgboost-feedstock#222 (comment)

So this appears to be working as intended

[trivialfis]

I find the behavior confusing when I work on this. I use everything from conda including cmake, compilers, and rmm. I have seen cases where cmake ignores cccl from conda env and jumps to system ctk. I have also seen it using the one from rmm (env/include/rapids/) without any configuration. I didn't try to debug it at the time, but this does give me a feeling that the configuration is fragile.

[trivialfis]

Merging, we can continue the discussion if the issue ever comes up again.

[trivialfis]

@jakirkham Should we backport it to the 3.0 branch?

[jakirkham]

Yes let's do that

Have submitted PR: #11354