Unconstrained Model Merging for Enhanced LLM Reasoning