Controlled Low-Rank Adaptation with Subspace Regularization for Continued Training on Large Language Models