[REVIEW] Deprecate CNMeM

[harrism]

This PR deprecates RMM. Depends on #449 ✔️ and #457 ✔️. Note this PR will be much simpler once those are merged.

Closes #407
Contributes to #459

• Removes the dependency on cnmem memory resources in RMM Python. Pool mode switched to pool_memory_resource
• removes cnmem from all tests
• removes cnmem from all benchmarks
• adds C++ deprecation keywords to cnmem_memory_resource and cnmem_managed_memory_resource
• Updates the readme to mention that cnmem resources are deprecated
• adds C++ deprecation keywords to get_default_resource and set_default_resource and updates readme.MD to describe the new per-device default resource APIs.
• replaces usage of C++ get_default_resource with get_current_device_resource
• Makes C++ get_default_resource and set_default_resource a wrapper around get/set_current_device_resource
• Updates Python rmm.reinitialize() to create a resource per specified device
• Updates Python API to expose current device and per-device resource APIs rather than default resource APIs
• Updates readme

This PR does not touch cmake -- that will be left for the release where we remove cnmem.

Need to follow this up with PRs for various RAPIDS repos to not use cnmem resources
in test fixtures and benchmarks.

[harrism]

@kkraus14 I'm going to add some more changes -- expose per-device resource APIs in Python since we want to deprecate the old ones (get/set_default_resource) along with deprecating cnmem...

[kkraus14]

@kkraus14 I'm going to add some more changes -- expose per-device resource APIs in Python since we want to deprecate the old ones (get/set_default_resource) along with deprecating cnmem...

Can you mark as WIP unless you're planning on doing those in a follow up PR 😄

More changes incoming

[harrism]

Changed my mind, should be a follow up PR.

[kkraus14]

I think we can remove the brackets here as this just unnecessarily creates an intermediate Python list instead of using a generator expression:

Suggested change

	[each_mr.c_obj.get() is not NULL
	for each_mr in _per_device_mrs.values()]
	each_mr.c_obj.get() is not NULL
	for each_mr in _per_device_mrs.values()

[harrism]

closures inside cpdef functions not yet supported

[harrism]

Invalid types for 'is_not' (Python object, void *)

[kkraus14]

well then, looks like we stay with the Python list for now 😅

[harrism]

I can't get isort to pass -- running it locally on the failing file doesn't do anything.

[kkraus14]

Hmm, historically we've allowed importing rmm / cudf / etc. on machines without a GPU. With the change to RMM initialization we now call a cudaGetDevice that will throw if there's no GPU available.