You are setting up CI/CD for an LLM-powered application. Explain what makes this fundamentally different from traditional ML CI/CD and from software CI/CD. Cover: how to run automated evals in CI (LLM-as-judge, regression suites), how to manage prompt versioning as code, how to handle non-deterministic test outputs in CI pipelines, how to gate deployments on eval metrics rather than just test pass/fail, how to do canary deployments for prompt changes, and what rollback looks like when a prompt change causes quality regression in production.