TensorLLM: Enhancing Reasoning and Efficiency in Large Language Models through Multi-Head Attention Compression and Tensorisation