Add LLM cache for query expansion and reranking

[aurexav]

Context
QMD caches LLM outputs to keep query expansion and reranking consistent and inexpensive on repeated queries.

Goal
Reduce latency and cost for repeated queries while keeping outputs consistent.

Scope

• Cache expansion and rerank results keyed by query, model, and version.
• Define cache invalidation rules and a retention policy.

Acceptance Criteria

• Repeated queries avoid redundant LLM calls without correctness regressions.

References

• https://github.com/tobi/qmd