Quantization-Aware and Tensor-Compressed Training of Transformers for Natural Language Understanding