Training Language Model Agents to Find Vulnerabilities with CTF-Dojo