feat(config, llm): Add a vllm provider

[rxgrant]

For vLLM compatibility, the openai provider has the Bearer token auth, but it speaks a Responses dialect that vLLM does not serve. The llamacpp provider speaks the Chat Completions dialect that vLLM serves, but it sends no auth and it strips the vendor prefix from the model id. We need a new provider for vLLM.

Users can now point jp at a self-hosted vLLM server by setting a model's provider to vllm. vLLM speaks the same OpenAI-compatible /v1/chat/completions dialect as llama.cpp, so it supports streaming chat, tool calls, structured output, and thinking control the same way. Model listings come from GET /v1/models, and the reported max_model_len becomes the model's context window; the model name keeps its full id (e.g. Qwen/Qwen3-8B) since vLLM only accepts that form in requests.

providers.llm.vllm configures the new provider, with api_key_env defaulting to VLLM_API_KEY and base_url defaulting to http://127.0.0.1:8000:

[providers.llm.vllm]
api_key_env = "VLLM_API_KEY"
base_url = "http://127.0.0.1:8000"

To support this without duplicating llama.cpp's SSE parsing and event conversion, the shared OpenAI-compatible message/tool conversion and stream assembly logic moves from provider/llamacpp.rs into provider/openai_compat.rs, where both providers now call it. This is behavior-preserving for llama.cpp.

Closes: #0khs4b1

[rxgrant]

rebased