On-Demand Multi-Task Sparsity for Efficient Large-Model Deployment on Edge Devices