APOLLO: An Optimized Training Approach for Long-form Numerical Reasoning