research SpecBridge Natural Language to Lean Specifications Cross-Architecture Operator Testing Differential testing for GPU/NPU neural network operators MoE Inference Acceleration Adaptive acceleration for Mixture-of-Experts large language models