Disperse-Then-Merge: Pushing the Limits of Instruction Tuning via Alignment Tax Reduction