UniMod1K: Towards a More Universal Large-Scale Dataset and Benchmark for Multi-modal Learning