[doc] Enhance documentation for distributed training.

Currently, we have some high-level documents for various interfaces, such as Dask and Spark. But the documentation around actual XGBoost components is scant. It would be great if we could have a more coherent document specifically for distributed training, to aid third-party integration and customization.

This is particularly important, as existing distributed frameworks are no longer sufficient for training large models when we want to max out hardware capacity.

related:
- https://github.com/kubeflow/trainer/pull/3118

The issue names no files or tests; start by reviewing the existing Dask and Spark interface documents and the XGBoost distributed-training components. Done means a coherent XGBoost-focused distributed-training document that explains third-party integration and customization.

Written by the indexing model from the issue text.