Where Do LLMs Still Struggle? An In-Depth Analysis of Code Generation Benchmarks