[FEA] RAFT benchmarks should allow memory type to be specified at runtime

[cjnolet]

As an example, the MemoryType for the dataset in the CAGRA benchmarks is currently hard-coded to HostMmap here, which means anyone using the conda package to benchmark is going to end up using that (and unknowingly taking a perf hit as a result). We should allow that to be specified at runtime.