SlopCodeBench: Benchmarking How Coding Agents Degrade over Long-Horizon Tasks