APT-LLM: Exploiting Arbitrary-Precision Tensor Core Computing for LLM Acceleration