Add LLM response cache with SQLite backend

[bug-ops]

Parent: #740 (P2)

Problem

Identical or near-identical queries re-invoke LLM, wasting tokens and latency.

Solution

• SQLite-backed response cache keyed by content hash (SHA-256 of messages + model + params)
• Configurable TTL (default: 1 hour)
• Config: llm.response_cache.enabled (bool), llm.response_cache.ttl_secs (u64)
• Periodic cleanup of expired entries

Affected crates

• zeph-memory (cache storage)
• zeph-llm (cache lookup/store around provider calls)

Acceptance criteria

• Cache hit returns stored response without LLM call
• TTL-based expiry
• Cache bypass option per request
• Periodic cleanup