[BUG] ANN_BENCH: incorrect output index type in multi-gpu CAGRA

All multi-gpu implementations in ANN_BENCH ignore the benchmark index type algo_base::index_type, which is int64_t:

• https://github.com/rapidsai/cuvs/blob/73e4fbcb981a67048a7f63b39416a87cc878c1ba/cpp/bench/ann/src/cuvs/cuvs_mg_cagra_wrapper.h#L161-L162
• https://github.com/rapidsai/cuvs/blob/73e4fbcb981a67048a7f63b39416a87cc878c1ba/cpp/bench/ann/src/cuvs/cuvs_mg_ivf_flat_wrapper.h#L134-L135
• https://github.com/rapidsai/cuvs/blob/73e4fbcb981a67048a7f63b39416a87cc878c1ba/cpp/bench/ann/src/cuvs/cuvs_mg_ivf_pq_wrapper.h#L131-L132

IVF-Flat and IVF-PQ index types coincide with the benchmark index type and thus show correct results.
CAGRA index type is always instantiated as uint32_t. As a result, int64_t array is casted to int32_t, and the benchmark interprets the algorithm output incorrectly.

We either need to map the outputs correctly as it's done in IVF-PQ benchmark wrapper (which adds extra overhead if the types do not match and puts the index at disadvantage in the benchmarks) or adjust the implementation of multi-gpu indexes to decouple index-index type from the search output index type as it was done for CAGRA in https://github.com/rapidsai/cuvs/pull/769 .

Start with the multi-GPU wrappers at cpp/bench/ann/src/cuvs/cuvs_mg_cagra_wrapper.h:161-162, cuvs_mg_ivf_flat_wrapper.h:134-135, and cuvs_mg_ivf_pq_wrapper.h:131-132, then compare their output handling with cuvs_ivf_pq_wrapper.h:181-222 and cuvs_cagra_wrapper.h:368-369. Decide whether to map output types or decouple the index and search output types, then run ANN_BENCH to confirm multi-GPU CAGRA results use the benchmark's int64_t index type.