SciSafeEval: A Comprehensive Benchmark for Safety Alignment of Large Language Models in Scientific Tasks