[FEA] Provide an escape hatch to libraries that operate cudf aware

[galipremsagar]

Is your feature request related to a problem? Please describe.
cudf.pandas acts a proxy layer wrapped around cudf or pandas objects. There are libraries like xgboost and cuml that choose different code-paths or handle the objects differently depending on if the object is residing on GPU memory or CPU memory. A classic example is xgboost:

https://github.com/dmlc/xgboost/blob/c0f1f568996a059abfa799d2fa4b6433de894a09/python-package/xgboost/sklearn.py#L1578-L1603

When cudf.pandas objects go through these code-flows, it becomes harder and more compute/transfer intensive to move cudf.pandas objects to CPU memory. The reason being cudf.pandas objects will never pass is_cudf_.. checks and rather will always pass is_pandas_... checks. Thus cudf.pandas objects end up being forced to take CPU code paths. This can be avoided if we provide a public API that can be called on cudf.pandas proxy objects to return a GPU object.

I recently made a fix to xgboost as a stop-gap measure to make cudf.pandas proxy objects extract GPU objects: dmlc/xgboost#11014

Rather than accessing private attributes, it would be better to have a public API that will return the GPU object.