Apple study finds "a fundamental scaling limitation" in LLM reasoning models