Haystack 3.3.0 speeds up SentenceWindowRetriever and updates BM25 behavior

Haystack 3.3.0 reduces SentenceWindowRetriever's Document Store lookups to one per run or run_async call, rather than one per retrieved document, while keeping outputs unchanged. That means existing pipelines can see lower latency and less Document Store load without code changes. InMemoryDocumentStore.bm25_retrieval and InMemoryBM25Retriever using BM25L or BM25Plus now return only documents containing a query term, potentially fewer than top_k or none. Documents that match may also score lower when query terms are absent from them, so teams using fixed score thresholds should reassess those thresholds; BM25Okapi is unaffected. Separately, the default BM25 tokenization regex now splits Chinese, Japanese and Korean characters individually and NFC-normalizes text, allowing bare-term queries to match within unspaced text. Haystack also raises its minimum anyio requirement to 4.14.2 because earlier versions are affected by CVE-2026-63374.