[doc][EM] Add a brief introduction to NUMA.

[trivialfis]

Add a utility to help set the CPU affinity.

[pentschev]

Why? https://github.com/rapidsai/dask-cuda/blob/6c3223d7f82f4b2e7eb209a8485b1e32cb5357a7/dask_cuda/utils.py#L752-L775

[trivialfis]

Thank you for pointing it out. Got a bit too used to cudart.

[trivialfis]

Thinking again, I need this cudart version as XGBoost should prefer the CUDA device enumeration instead of the nvml device enumeration.

[trivialfis]

@pentschev I have simplified the utility using nvml and changed the package name in the documentation.

[hcho3]

Let's document that set_device_cpu_affinity is not available on other OSes like Windows.

[trivialfis]

Update:

• Documented that this is Linux-only.
• Use CUDA for device enumeration. We obtain the device ordinal and corresponding UUID from CUDA, then use the UUID to get the affinity. This way, we can honor the CUDA_VISIBLE_DEVICES environment variable.
• Manually tested on an NVL system.

[trivialfis]

Hold this PR for now. I think the CPU affinity alone is not sufficient when the memory is under pressure.

[trivialfis]

Expanded the document for using numactl, and removed all the utilities. Removing the utilities since CPU affinity alone is not sufficient when memory is under pressure. I don't want to maintain specialized code that is not effective.

[Copilot]

Pull Request Overview

This PR enhances the External Memory tutorial by introducing a section on NUMA configuration and updates the Python demos to reference the new cuda-python package and the NUMA guidance.

• Added a table of contents and a new NUMA section explaining how to set CPU and memory affinity.
• Updated demo scripts to switch from python-cuda to cuda-python, adjust the cudart import, and reference the NUMA tutorial.
• Minor history update in the release notes to include support for the Grace Blackwell decompression engine.

Reviewed Changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
doc/tutorials/external_memory.rst	Added TOC and a detailed NUMA section with examples for numactl.
demo/guide-python/external_memory.py	Swapped python-cuda to cuda-python, updated import, added NUMA note.
demo/guide-python/distributed_extmem_basic.py	Same updates as above: package name, import path, and NUMA reference.

Comments suppressed due to low confidence (3)

doc/tutorials/external_memory.rst:297

• [nitpick] Consider adding a brief note or link on installing numactl (e.g., via apt-get install numactl), so readers know how to obtain the tool before using it.

    numactl --membind=${NODEID} --cpunodebind=${NODEID} ./myapp

demo/guide-python/external_memory.py:50

• [nitpick] The device_mem_total function is duplicated across demos; consider extracting it into a shared utility module to reduce repetition.

    import cuda.bindings.runtime as cudart

demo/guide-python/distributed_extmem_basic.py:44

• [nitpick] Same helper appears here; extracting the GPU memory query into a common helper would improve consistency and reduce maintenance overhead.

    import cuda.bindings.runtime as cudart