dask-cuda uses pandas for some tests, but the main reason for the pinning is that it is inherited from RAPIDS libraries (mainly cudf) that do not yet support pandas 2.0 and are the primary use case for dask-cuda. However, there is no reason dask-cuda cannot be used in other contexts, so relaxing this constraint makes sense.
In order to unblock cudf's pandas 2.0 testing I'm going to merge this, but I'll continue the discussion offline about an appropriate pandas lower bound (or if there needs to be one here at all).
dask-cuda uses pandas for some tests, but the main reason for the pinning is that it is inherited from RAPIDS libraries (mainly cudf) that do not yet support pandas 2.0 and are the primary use case for dask-cuda. However, there is no reason dask-cuda cannot be used in other contexts, so relaxing this constraint makes sense.
Resolves #1306