CoreInfer: Accelerating Large Language Model Inference with Semantics-Inspired Adaptive Sparse Activation