Checkpointing support for transformer type models

[zhenghh04]

In this PR, we addressed the issue that people have to manually input layer parameters and optimization groups in the checkpointing. #248

• Now, people are allowed to input just "vocab_size, hidden_size, ffn_hidden_size, num_layers", the layer parameters and optimization groups are calculated internally.
• We also added option for specifying zero_stage. The default value is -1, where no zero is used. Data parallel instance 0 will checkpoint the whole data.
• We added several YAML configuration files for Llama transformer models.
• We allow people to specify the datatype for writing data.

[zhenghh04]

@hariharan-devarajan This PR is ready for review.