LLMClient: OLLAMA_HOST/OPENAI_BASE_URL ignored, keyless local servers unusable, explicit model can be overridden #184

ebuild/eos_ai/llm_integration.py on master:

• ignores OLLAMA_HOST (a schemeless host:port, as Ollama documents it), OPENAI_BASE_URL and the model env vars;
• is_available() requires an API key for custom, so a keyless local OpenAI-compatible server (vLLM, llama.cpp) is never used, and auto() needs EOS_LLM_API_KEY as well as EOS_LLM_URL;
• sends Authorization: Bearer with an empty key;
• treats a response with no choices as an empty success instead of a failure;
• model defaults to llama3 for every provider, so LLMClient(provider='openai') asks OpenAI for llama3.

PR #127 addresses all of these.