NonGEMM Bench: Understanding the Performance Horizon of the Latest ML Workloads with NonGEMM Workloads