Improve error handling of embedding responses

Even a successful HTTP request to an embedding model could contain an error response. The LLM Application should try parsing it and use it to provide a more helpful error message.