Adaptive Orchestration for Large‐Scale Inference on Heterogeneous Accelerator Systems: Balancing Cost, Performance, and Resilience | Synapse