EdgeBERT: Sentence-Level Energy Optimizations for Latency-Aware Multi-Task NLP Inference