CAGRA - runtime dispatch of distance functions

[achirkin]

Move the TEAM_SIZE and DATASET_BLOCK_DIM template parameters of CAGRA search kernels (single_cta and multi_cta versions) to runtime by introducing a switch-case statement inside the compute distance components (hence the name: runtime dispatch of the distance implementation). As a result:

• 4x less search kernel instances against branch-24.10, but each instance is slightly larger
• There's a small performance penalty due to larger kernel code (and the register usage is bounded by the most register-expensive branch)
• [implementation limitation] only one DATASET_BLOCK_DIM per TEAM_SIZE is possible, i.e. the team size defines both (though this is how it currently is anyway).

[achirkin]

This is an alternative to #296 for reducing the binary size with minimal changes to the code. Here's how they compare:

enh-cagra-runtime-distance-dispatch (#324)

performance: up to 12% QPS slowdown on deep-100M and 5% QPS slowdown on wiki-all
libcuvs.so size: 791 -> 657 MB
parallel build time in CI (NB: caching is not disabled):
- total: 709 seconds
- neighbors/cagra_search_xxx: 9 min
- neighbors/detail/cagra_xxx (kernels): 2-3 mins

Sheets: https://docs.google.com/spreadsheets/d/1sm-Qs7upa1JP706TnhX7_M2PFj57ZBNUIRk4F44ZvO0/edit?gid=347774484#gid=347774484

enh-cagra-separable-compilation (#296)

performance: up to 13% QPS slowdown on deep-100M and up to 8% QPS slowdown on wiki-all (low-itopk CAGRA-Q cases being the worst, with others seemingly slightly better than (#324)
libcuvs.so size: 791 -> 624 MB
parallel build time in CI (NB: caching is not disabled):
- total: 478 seconds
- neighbors/cagra_search_xxx: 4 min
- neighbors/detail/cagra_xxx (kernels): 1-3 mins
- neighbors/detail/compute_distance_xxx: 50 sec (many of them)

Sheets: https://docs.google.com/spreadsheets/d/1b_yTRcWE9K8SpEVGhzW_UZa2pTLibFSF8CFc2W3bJqM/edit?gid=347774484#gid=347774484

[achirkin]

To sum it up:

• the QPS of this PR is actually slightly better due to some VPQ cases being particularly bad in CAGRA - separable compilation for distance computation #296
• this PR reduces the compile size almost as good as CAGRA - separable compilation for distance computation #296, but kernels are still large; this means adding Persistent CAGRA kernel #215 on top of this PR will have larger impact on binary size, because it duplicates the single-cta kernel (approx 60-70MB)
• In contrast to CAGRA - separable compilation for distance computation #296, this PR does not reduce the compile time
• This PR introduces minimal changes, which is much easier for review
• CAGRA - separable compilation for distance computation #296 adds extra functionality (multi-kernel CAGRA-Q) and (:in my opinion:) makes it easier to do further changes (e.g. flexible code generation for instances and more team_size, dataset_block_dim combinations can be supported).

[achirkin]

Performance update:
After a few optimizations for the worst cases #296, its performance is generally better than this PR on deep-100M dataset (dim = 96) and slightly worse on wiki-all dataset (dim = 768, only VPQ distance tested due to the index size). In the worst case, #296 leads to QPS drop of 6-7%, with plenty of cases actually getting a decent speedup (against both this PR and base-24.10). The caveat here is that some optimizations in #296 are potentially transferable to the base branch, so this benchmark is not entirely apples-to-apples anymore.

Latest benchmarks (includes k = 10 and k = 100 for both datasets):
https://docs.google.com/spreadsheets/d/191f1sYsAUwPidncV4xDzgNtOhKx9sp6jJFvOfpB-nxs