Refactor pylibcugraph to avoid maintaining unnecessary pointers to weights, edge IDs, others

[rlratzel]

There's currently code in PLC for taking data provided by Python users to be used as edge weights and/or edge IDs (and possibly others) and creating device arrays managed by RMM for use by libcugraph. This data is kept as a member of the internal _GPUGraph class and occasionally referenced later to see if the graph was created with edge weights or IDs, but does not appear to be used anywhere else.

This task is to investigate if there's an alternate approach that avoids keeping the pointers to the data and making the change if possible.