Add MemoryResourceConfig to cudf-polars config

[TomAugspurger]

Description

This adds a new configuration option to cudf-polars' config, allowing users to specify the memory resource to create by default.

Currently, users either get the default behavior (typically a managed memory resource) or can pass in a concrete memory resource (see the new docs added in this PR).

In some cases, you might want to pass in a description of the memory resource to use:

1. In our unit tests, we might want to specify a CudaAsyncMemoryResource with a relatively small initial pool size
2. In a distributed environment, you can't pass around concrete memory resource objects (which can't be serialized)

This is strictly more flexible than the current option of setting POLARS_GPU_ENABLE_CUDA_MANAGED_MEMORY. By setting POLARS_GPU_ENABLE_CUDA_MANAGED_MEMORY=0 you'll get a CudaAsyncMemoryResource with its default initial_pool_size and release_threshold. With this system, you can set

CUDF_POLARS__MEMORY_RESOURCE_CONFIG__QUALNAME="rmm.mr.CudaAsyncMemoryResource"

to get the same thing, or

CUDF_POLARS__MEMORY_RESOURCE_CONFIG__QUALNAME="rmm.mr.CudaAsyncMemoryResource"
CUDF_POLARS__MEMORY_RESOURCE_CONFIG__OPTIONS='{"initial_pool_size": 256, "release_threshold": 256}"'

to configure the pool.

I'd recommend deprecating POLARS_GPU_ENABLE_CUDA_MANAGED_MEMORY to reduce the number of ways this can be configured, though we should take our time with that.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[TomAugspurger]

Are people OK with "qualname" here? I want to avoid locking us to MRs that happen to be defined in rmm.mr.

[TomAugspurger]

The only other feature I might want to add here is expanding the MemoryResourceConfig to handle "nested" memory resource configurations. Right now, you can't create an RMM memory resource that needs to configure the inner memory resource. I'm not sure whether that's worth doing here or not.

For example, to express our default memory resource:

            mr = rmm.mr.PrefetchResourceAdaptor(
                rmm.mr.PoolMemoryResource(
                    rmm.mr.ManagedMemoryResource(),
                    initial_pool_size=free_memory,
                )
            )

Aside from free_memory itself being dynamic, this could be expressed as something like:

{
    "qualname": "rmm.mr.PrefetchResourceAdaptor",
    "options": {
        "upstream_mr": {
            "qualname": "rmm.mr.PoolMemoryResource",
            "options": {
                "upstream_mr": {
                    "qualname": "rmm.mr.ManagedMemoryResource",
                    "options": {
                        "initial_pool_size": 256
                    }
                }
            }
        }
    }
}

That relies on pattern matching a dict with {"qualname": ..., "options": ...} to mean "this is a memory resource config", which is probably sufficient.