Apple paper: “LRMs effectively overthink on simple problems, outperform LLMs on mid-complexity puzzles, and fail in the same exact way LLMs do on high-complexity tasks and puzzles”