Require rapidsmpf Pool with the rapidsmpf runtime

[TomAugspurger]

Description

This updates our config to require using the rapidsmpf CUDA Stream Pool with the rapidsmpf runtime. As we move pieces of the IR execution into native rapidsmpf IR nodes (e.g. rapidsmpf's read_parquet), which necessarily have to occur on a stream from the rapidsmpf stream pool, it seems like the complexity of mixing multiple stream sources (mostly missing a necessary CUDA stream synchronization) is not worth any potential benefits.

Closes https://github.com/rapidsai/cudf/issues/20756