bb25 v0.4.0 upgrades hybrid retrieval core

This is the kind of retrieval release that moves beyond “just blend BM25 and vectors” into disciplined ranking engineering.

• –BlockMaxIndex (BMW) and upper-bound pruning target real latency wins by skipping low-potential candidates during retrieval.
• –Multi-head attention in log-odds space is a meaningful step up from static weighted sums, especially for query-dependent fusion behavior.
• –Platt and isotonic calibration matter because calibrated probabilities are easier to threshold, monitor, and reason about in production.
• –Temporal Bayesian decay is practical for freshness-sensitive workloads like news, logs, and rapidly changing knowledge bases.
• –Community traction around bb25-style calibration and hybrid search suggests this is landing in an active RAG/search optimization niche.