LongCodeBench: Evaluating Coding LLMs at 1M Context Windows