Optimizing Dynamic-Shape Neural Networks on Accelerators via On-the-Fly Micro-Kernel Polymerization | Synapse