Pushing the limits of long-context LLM training for 1M-token+ medical records