feat: add --server-port option to connect to external vLLM server

[CharlesR-W]

Summary

Adds an optional --server-port CLI argument that allows connecting to an external vLLM server via its OpenAI-compatible API, instead of loading the model in-process.

This is useful when:

• Sharing a single model server across multiple delphi runs
• Managing GPU memory separately from the delphi process
• Running the model on a different machine

Usage

# Start vLLM server separately
vllm serve meta-llama/Meta-Llama-3.1-70B-Instruct-AWQ-INT4 --port 8000

# Run delphi with server mode
python -m delphi --server-port 8000 ...

Default behavior is unchanged (server_port=None loads model locally as before).

Changes

• delphi/clients/offline.py: Add server_port parameter, _generate_server() method for OpenAI API calls
• delphi/config.py: Add server_port config field
• delphi/__main__.py: Pass server_port to Offline client
• pyproject.toml: Add openai>=1.0.0 dependency

🤖 Generated with Claude Code

[CLAassistant]

All committers have signed the CLA.