This PR adds support for using an external vLLM server with the pipeline, enabling significant performance improvements by keeping the model loaded between runs.

Changes:
- Add vllm_server_manager.py: Standalone script to manage a persistent vLLM server with health monitoring and automatic restart capabilities
- Add --vllm-url flag to pipeline.py: Allows connecting to an external vLLM server instead of starting one internally
- Add model verification: Ensures the external server has the correct model loaded before processing
- Add robust retry logic: Handles connection failures with exponential backoff and health checks

Benefits:
- Save 30-60+ seconds per pipeline run by avoiding model reloading
- Run multiple pipeline instances against the same server
- Better resource utilization with persistent GPU memory allocation
- Automatic server restart on crashes with configurable retry logic
- Clean separation of server infrastructure from processing logic

Usage:
1. Start the server manager: python -m olmocr.vllm_server_manager
2. Run pipeline with external server: python -m olmocr.pipeline workspace --vllm-url http://localhost:30024 --pdfs *.pdf

The implementation is fully backward compatible - the pipeline works exactly as before if --vllm-url is not provided.