Low-Rank and Sparse Model Merging for Multi-Lingual Speech Recognition and Translation