AbsPyramid: Benchmarking the Abstraction Ability of Language Models with a Unified Entailment Graph