perf: cache PreparedQuery per stream, skip double-normalize in dedupe (#282)

Scoring hot path (_normalize_score_dedupe) re-tokenized the same
ranking_query ~240x per stream: once per item for local_relevance,
plus ~5x per item across snippet windows. Query tokens are immutable
within a stream, so compute them once as relevance.PreparedQuery and
thread through signals.annotate_stream and snippet.extract_best_snippet.

dedupe._PreparedText called normalize_text twice: once in __init__ and
again via get_ngrams. Factor out _ngrams_of_normalized so the prepared
path skips the redundant pass while get_ngrams keeps its public contract.

Behavior unchanged.

skills/last30days/scripts/lib/snippet.py

@@ -26,7 +26,7 @@ def _windows(words: list[str], size: int, overlap: int) -> list[str]:
 
 def extract_best_snippet(
     item: schema.SourceItem,
-    ranking_query: str,
+    ranking_query: "str | relevance.PreparedQuery",
     max_words: int = 120,
 ) -> str:
     """Prefer existing snippets, else extract the best matching evidence window."""
@@ -43,8 +43,9 @@ def extract_best_snippet(
     if not candidates:
         return _truncate_words(body, max_words)
 
+    prepared_query = ranking_query if isinstance(ranking_query, relevance.PreparedQuery) else relevance.PreparedQuery(ranking_query)
     best = max(
         candidates,
-        key=lambda candidate: relevance.token_overlap_relevance(ranking_query, candidate),
+        key=lambda candidate: relevance.token_overlap_relevance(prepared_query, candidate),
     )
     return _truncate_words(best, max_words)