Improve `device_resources_snmg` RMM pool robustness

[viclafargue]

Following #2866, the device_resources_snmg's set_memory_pool function correctly sets the memory pool on each GPU with RMM. Unfortunately, once this function has been called the global RMM memory resource is tied to the device_resources_snmg instance. Hence anything allocated after calling set_memory_pool should be released before the device_resources_snmg instance is released and would otherwise result in segfaults.

To remediate this issue the set_memory_pool function should be updated to use the new RMM memory resources (see #2866 (comment)).