Add note that --rmm-async only affects distributed scheduler.

[bdice]

Description

The PDS-H queries don't obey --rmm-async unless using the distributed scheduler. This adds a note to the docs.

Perhaps we should solve this differently, so that --rmm-async actually works for all schedulers?

A workaround is setting POLARS_GPU_ENABLE_CUDA_MANAGED_MEMORY=0.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[TomAugspurger]

Sorry for missing this earlier.

Perhaps we should solve this differently, so that --rmm-async actually works for all schedulers?

That'd be done at

cudf/python/cudf_polars/cudf_polars/experimental/benchmarks/utils.py

Lines 752 to 756 in dcf7eff

	engine = pl.GPUEngine(
	raise_on_fail=True,
	executor=run_config.executor,
	executor_options=executor_options,
	)

. It'd be a bit of logic to parse out the rmm-related options and construct the memory_resource argument to GPUEngine.

#20042 is also a bit related here, which adds another (slightly more flexible) way to configure RMM MRs.

I think that your change here is a good improvement and we can merge it as is. I can handle integrating these RMM options into that as part of #20042.

[bdice]

/merge