search: precompute intent once, reuse across all semantic_search calls

[donsummerwind]

Summary

Precompute LLM intent analysis once per search, then reuse across all semantic_search calls. Reduces LLM calls from 5 to 1 per search.

Changes

• SearchOptions gains precomputed_intent: Option<Arc<EnhancedQueryIntent>>
• search_handler calls analyze_intent() once before layered_semantic_search
• layered + 4 semantic_search calls all reuse the same intent
• Result: ~3-5x search speedup

Performance

Before	After
5× LLM calls (~16s)	1× LLM call (~3s warm, 7B cold ~8s)

Files changed

• cortex-mem-core/src/search/vector_engine.rs
• cortex-mem-core/src/types.rs
• cortex-mem-core/src/vector_store/qdrant.rs
• cortex-mem-service/src/handlers/filesystem.rs
• cortex-mem-service/src/handlers/search.rs
• cortex-mem-service/src/main.rs