Load a model into memory.

Pass {"model_name": "kokoro"} to target any registered engine. Only an absent model_name falls back to the default Qwen TTS backend, selected by the model_size query parameter, which is what callers did before model_name existed. A supplied-but-empty name is a bad request, not a request for the default — silently loading Qwen there is the very bug this endpoint had. Either way a TTS load goes through load_engine_model (VoiceLabs: one heavy model resident at a time).