Generalize array checking and remove pd.Index call in _get_partitions

[quasiben]

This PR does two things:

1. Generalizes the array checking using is_arraylike as well as the list check
2. Removes what I think is an unnecessary pd.Index call .

This fixes an issue for dask-cudf where users want to index a Dask Dataframe with a cudf/cupy object:

df.loc[cudf.Series([0,10,50], dtype=np.int32)]

The pd.Index call is quite old (#1913) and was originally written to handle indexing with lists. Locally I ran dask/dataframe/tests/test_indexing.py and all tests still pass as well as verifying manually that lists are still supported

• Closes #xxxx
• Tests added / passed
• Passes pre-commit run --all-files

cc @VibhuJawa @rjzamora

[VibhuJawa]

Thanks for working on this @quasiben , I expect to see some speed up for the GNN use cases when this lands.

CC: @eriknw for awareness too in relation to get_edge_data and get_node_data for PropertyGraphs.

[jrbourbeau]

Thanks @quasiben! Can we add a test for this?

[quasiben]

Thanks for the review @jrbourbeau -- added a test and found another spot where we needed another array-like call

[jrbourbeau]

My guess is the gpuCI failures are unrelated (saw them here yesterday #9635 (review)), though it'd be great if you could confirm

[quasiben]

@jrbourbeau , yes, they are unrelated. I filed #9639 to track the failures

[jrbourbeau]

Thanks @quasiben