Prior to this change, raft's host-side implementation of raft::neighbors::refine operation uses non-optimal OpenMP thread config by default, spawning as many threads as there are available cores, even if only one thread is used (per-query parallelism with batch size one).
This change fixes that and adds a few optimizations alongside:

• Use the number of threads = min(cores, queries)
• Use templates to push the metric-type condition outside the distance computation loop (this should also make it easier to implement new metrics later)
• Force tree-optimize compilation flag in the hopes compiles does the vectorization
• Split out the host implementation in separate files to be able to compile it without nvcc.