Provider for OpenAI-compatible local servers (llama.cpp, vLLM, LM Studio) #10

Problem
Only Ollama is supported for live runs. llama.cpp's llama-server, vLLM, and LM Studio all expose an OpenAI-compatible /v1/chat/completions endpoint locally. One provider would cover all three, and it would also make it possible to compare runtimes against each other.

Expected behavior

id: llamacpp-q4
provider: openai_compatible
base_url: http://localhost:8080/v1
model: local-model
options: {temperature: 0, seed: 42}

The provider sends messages: [{"role": "user", "content": case.input}] (with an optional system), and reads choices[0].message.content. It uses the standard library only, like the Ollama provider.

Scope / where
src/modeldiff/providers.py, CONFIG_SCHEMA in config.py, build_provider, and the README section on local models. See "Adding a provider" in docs/architecture.md.

Acceptance criteria

• No API key is required. If an optional api_key_env is supported, it reads the key from an environment variable, and the key never appears in result files.
• Connection errors, HTTP errors, and malformed responses become ProviderError with a clear message.
• Usage counts from the response are stored in provider_metadata.

Tests
Mocked urlopen tests modelled on the Ollama tests in tests/test_config_and_providers.py. No running server is needed.

Definition of done
Tests pass, the README is updated, and there is an example config in examples/.