Enhancing checkpoint support for transformer type of workloads

[zhenghh04]

Currently, when people want to do checkpointing write for transformer type workloads, one has to input the layer_parameters and optimization groups. All these parameters actually can be derived from the higher level parameters such as hidden dimension, ffn size, vocab_size, etc. Asking users to directly input layer_parameters and optimization groups in the configure file is not so good. We can hide all the lower level details here.