[FEA] Multi-GPU wrappers assume 64-bit indexing types

[cjnolet]

We should make this more flexible. @achirkin has been doing some benchmarks and finding the build and search times take a perf hit as a result of this.

Artem to provide exact perf degredation here. We discussed offline the potential for us to separate the vertex ids from the graph in order to be more flexible and not suffer this perf regression.

[viclafargue]

Please correct me if I'm wrong, but I believe there is no performance or accuracy degradation. At the time the SNMG API was created, the CAGRA index only supported building and searching with uint32_t indices. However, the benchmarking code expects int64_t indices for result verification.

Since this is a benchmark, converting the neighbor indices after the search would have falsified the benchmark results by introducing a transformation not present during the actual search process. To avoid that, I simply casted the array pointer (without modifying its content), which, of course, caused the verification accuracy to drop to zero.

CAGRA has since been updated and now supports returning int64_t indices during search, even when the index is built with uint32_t. This PR exposes the updated search function in the SNMG API and updates the benchmark code to use it, resolving the verification issue.