[FEA] CAGRA follow up

[tfeher]

#1375 adds CAGRA to RAFT as an experimental feature. Here is a list of issues that shall be addressed as a follow up:

Problems:

• CAGRA binary size too large. #1459
• CAGRA int8 / uint8 distance scaling #1457

Features:

• CAGRA support arbitrary dim (number of features) #1458
• support different metrics
• NN descent for index building
• add / remove vectors
• add python API

Tech debt:

• CAGRA multi-cta solver accuracy #1575
• consolidate top-k implementations: replace CAGRAS's top-k with RAFT top-k methods or add CAGRA's variant to raft top-k algorithms ([FEA] Consolidate CAGRA's Top-k implementations #1603)
• Compose CAGRA from RAFT primitives/helpers #1604
• add developer docs
• check whether we could use cuCollections for hashmap https://github.com/rapidsai/raft/pull/1375/files#r1159048450
• CAGRA reconcile multi gpu prune with RAFT multi GPU paradigm #1460
• CAGRA use rmm allocator and fix stream ordering in prune #1461

Improvements:

• [FEA] CAGRA avoid unnecessary device memory copies of dataset / knn graph #1479
• Re-introduce CAGRA template instantiations to reduce compile time  #1443