How TRM Labs Scaled Security with Self-Improving AI Vulnerability Agents