Text this: ST‐LoRA: SVD‐Guided Sparse Low‐Rank Adaptation With Trainable Masks for Large Language and Vision Models