Either deprecate or test support for URI inputs like s3.

[trivialfis]

Currently, the support for object storage like s3 and hdfs are "best-effort", meaning we write code using dmlc::Stream most of the time but with an exception in external memory support due to the use of mmap. We have no test for related features on our CI. I suggest that we deprecate the use of these features inside XGBoost and let third-party libraries handle them. For instance, to load a model in Python, one can first use a different library to fetch the model from s3 then use Booster to load from a byte buffer. This way, we can limit the scope of the code base to be more aligned with machine learning and have less concern with system programming.