Fix benchmark regressions from #1846

[achirkin]

Fix a few issues introduced to the benchmark executables limiting the functionality:

• Reinstate the dropped support for CAGRA-Q (compression_* parameters)
• Avoid multiple copies of the data and the graph on device in various code paths
• Avoid copying the data if it is device-accessible
• Make logical merge reuse one buffer slot per split and free it before reallocating, instead of appending new split copies on every set_search_dataset call.
• Make multi-GPU CAGRA benchmarks use the padded index and dataset view, which CAGRA search requires.
• Revert copy() to the shallow copy-constructor form used before, so all shared state is forwarded to per-thread wrappers.

Library changes (non-breaking):

• Modified the implementation of cuvs::make_vpq_dataset to bring the ability to quantize from host data without copying the whole dataset and to quantize from fp16 data (the relevant symbols were hidden).

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[achirkin]

/ok to test

[achirkin]

/ok to test

[cjnolet]

@achirkin its too late for anything to be merged into 26.08. You'll need to retarget to main.

[achirkin]

Sure. I'll just finish it up today and let you decide whether to re-target and whether to optionally backport the fixes later if necessary.

[tfeher]

Thanks Artem, most of it looks good, but I have some questions.

Accidentally approved.