Simplifying Momentum-based Positive-definite Submanifold Optimization with Applications to Deep Learning