Use rmm allocator in CAGRA prune (#1503)

This PR modifies the CAGRA prune function to use rmm allocator instead of a custom memory allocator. And also removes the multi-GPU functionality that is not suitable for the [RAFT policy](https://docs.rapids.ai/api/raft/stable/developer_guide/#multi-gpu).

rel: #1461

Authors:
  - tsuki (https://github.com/enp1s0)
  - Tamas Bela Feher (https://github.com/tfeher)

Approvers:
  - Tamas Bela Feher (https://github.com/tfeher)

URL: #1503