TensorFlow 2.3 (v1 compat), Python 3.8, CUDA 10.2, CUDNN 7, AVX2 MKL XLA

Build from TensorFlow 2.3 with v1 compatible mode. The following bazel flags are used:

--config=v1 --config=xla --config=nohdfs --config=noaws --config=nogcp  --config=numa --config=mkl --config=cuda --config=opt 

cuda compute compatibility: 6.0,6.1,7.0

tensorflow/tensorflow#31481 was reverted to workaround tensorflow/tensorflow#41539