Improving Chain-of-Thought Reasoning via Quasi-Symbolic Abstractions