feat: bench cost reporter — per-run and per-task cost

[jpleva91]

Problem

LLMint has DeepSeek pricing but isn't wired into the bench pipeline.

Acceptance Criteria

• llmint cost <job-dir> computes per-task cost from token counts
• Uses model pricing from LLMint pricing table
• Outputs JSON for evolve pipeline integration

---

Generated by /forge cascade