Async LLM inference in CI: stop build workers blocking on slow jobs

Build workers on Buildkite were previously stuck for up to 35 seconds waiting on lengthy LLM inference jobs that summarized failed test outputs, causing significant delays and wasted compute resources. By implementing asynchronous inference through an AI gateway, workers can now submit jobs and get a unique ID, then poll later, freeing up resources and allowing other builds to proceed without waiting. This change is crucial for maintaining efficient CI/CD pipelines and reducing bottlenecks in a high-concurrency environment.