Modify tf bert-base model with frozen graph

[wangkl2]

• Modify the instruction for eval and infer, separately
• Enable padding for dGPU if the number of examples not a multiple of the batch size, to avoid performance degradation on dynamic shape.
• Add perf calculation for each step.
• Add the link to bert-base model with frozen graph in "language modelling" section in README. There is also another bert base inference link in "language translation" section but the performance with checkpoint is not as good as frozen graph.