Creates a runtime LLM (admin).

Creates a UI/API runtime LLM. The request body is the model configuration (display_name, base_model, inference_server, and any LiteLLM/provider fields) and is stored verbatim. The connection is probed before the model is saved; a failed probe rejects the create. With skip_test=true, the configuration is saved as not_probed and is unavailable for inference until a test of the saved model succeeds. Skipping does not bypass structural validation. A skipped configuration waits for a manual test while it is still not_probed, including after a restart. A FAILED manual test records probe_failed, which does put the configuration back into the startup re-probe population: once its endpoint recovers, a restart promotes it to ok and grants it by the default rules with no further admin action. display_name must be the name the model is served under whenever that name is a LiteLLM proxy routing key: the lite_llm_yaml block's top-level model_name when the config has a block, else base_model when inference_server is "litellm". A different display_name is rejected (400), because the model is registered with the proxy under that name. A model reached directly at its provider (inference_server anthropic, google, a vLLM endpoint, ...) with no block is registered under display_name itself, so its display_name is free. A block mux cannot read (hand-written YAML, or JSON declaring no top-level model_name, including model_list) is rejected. Admin only.