CAGRA search: int64_t indices in the API

[achirkin]

Add int64_t as the type of returned neighbors in cagra::search while keeping the index type unchanged.

Down the implementation, the PR introduces the OutputIndexT template parameter and handles it differently for each implementation:

• multi-kernel implementation already does the copy after search in a separate kernel, we change the index type during this copy.
• multi-cta kernel runs as before, followed by top-k; conversion of the index type happens in a separate map kernel only if a simple type cast wouldn't suffice (i.e. data sizes do not match).
• single-cta kernel must always run as a single kernel to maintain 'persistent' mode performance. I erase the index type in the kernel signature and pass the sizeof(OutputIndexT) as a tag in the pointer. Based on this tag, a proper output-writing function is dispatched in the epilogue of the single-cta kernel. By using this technique I avoid instantiating multiple kernels at the negligible runtime cost.

Note: we cannot simply instantiate cagra with index type int64_t, because this will make the graph twice larger and will reduce the search throughput.

To test the new search overloads, I turn AnnCagraTest::testCagra into a template and add just two instances to the *.cu files to avoid bloating the number of tests.

The new search overloads are used in the ann-bench wrappers, which slightly improves the reported QPS in some cases due to reducing the index conversion overhead.

[tfeher]

Thanks Artem, it looks good, just a few documentation issues. It is great that you could implement this without additional kernel instantiations.

[tfeher]

Thanks Artem for the update LGTM!

[tfeher]

/merge