Regret Analysis of Policy Optimization over Submanifolds for Linearly Constrained Online LQG