[FEATURE] llm matcher: inputs for multi-response semantic diff

Follow-up to #7733. Add inputs to the llm matcher so it compares several responses instead of one part.

matchers:
  - type: llm
    inputs: ["{{body_1}}", "{{body_2}}"]
    prompt: "Do these reveal whether one username exists and the other does not? Ignore tokens, timestamps, request IDs."
    expect: "yes"
    min-confidence: 0.85

Enables boolean-based blind SQLi, IDOR, and username enumeration without per-app tuning. New field plus a compose step before the prompt; the classify path is unchanged.