such as norm_train and to_static_train

2. fix some bugs in test_train_dy2static_python.sh

benchmark needs the field to set Global.print_batch_step=1. Check file "./test_tipc/benchmark_train.sh#L76" for details.

it will time out when turn on trt or mkldnn

the ips is unstable in the first 50 iter when training lite modle.

In benchmark_train mode, the config name is benchmark.txt. So the keyword of ShiTu should be model_name instead of the config name.

The benchmark requires threads num to be set 4 when DALI is used for best performance.
And the param num_workers is not worked before commit 80c10bb99503eeb349a5d676bd0af4908a74b425.