[FEA] Pre-filtering of nearest neighbors

[cjnolet]

Many end-to-end vector similarity search systems like Milvus and RedisAI VSS perform filtering of neighbors. For example, there might be indices removed from the database that need to be excluded from search, or only certain indices which have matched some search criteria might need to be included/excluded. In general, a boolean filter should be able to satisfy at least most of these cases but this issue is meant to drive discussion in addition to requesting a new feature.

Tagging some relevant folks @alexanderguzhva @liliu-z @achirkin @tfeher @benfred @wphicks

Tracking this feature's integration into RAFT's ANN algorithms:

• Brute-force
• Random Ball Cover
• IVF-Flat (Add sample filtering for ivf_flat. Filtering code refactoring and cleanup #1541)
• IVF-PQ (Introduce sample filtering to IVFPQ index search #1513)
• CAGRA Pre-filtering #1598

[Presburger]

Perhaps by doing post-filtering, the amount of data to be processed is much smaller compared to pre-filtering, and the computational overhead is actually not much compared to memory copy.