[BUG] Make cudf.pandas and cudf-polars compatible with RAPIDS_NO_INITIALIZE

[vyasr]

Describe the bug
Currently we support using the environment variables RAPIDS_NO_INITIALIZE and CUDF_NO_INITIALIZE to defer GPU initialization on import so that cudf will import successfully on machines without a GPU. However, this does not currently work with cudf.pandas because cudf.pandas tries to initialize a managed memory resource when it starts, leading to an error failing to find the CUDA driver.

Note that I would not expect this to occur with cudf-polars since cudf-polars only initializes the mr when the first collect call happens, not on import.

Steps/Code to reproduce bug
Install cudf on a machine without a GPU (probably easiest to test in a container without GPUs mounted) and run RAPIDS_NO_INITIALIZE=1 python -m cudf.pandas -c 'import pandas as pd'

Expected behavior
The above command should work. We can probably fix this by nesting the calls to pylibcudf.utils._is_concurrent_managed_access_supported inside a try-except block that can catch these failures and ignore them. We could also do some eager check for a GPU existing, but since the vast majority of practical use cases will have them I don't think LBYL is worthwhile.