[Orca] Support saving/loading trained TF model with h5/saveModel format

[sgwhat]

Description

1. Why the change?

• Current save and load do not support h5 and savemodel format, since they mainly retrieve model states.
• Users need the API to save model as h5 or savemodel format.

2. User API changes

For new added API:

• save model

  estimator.save(filepath, overwrite, include_optimizer, save_format, signatures, options)
  

• load model

  estimator.load(filepath, custom_objects, compile)
  

For original save/load API:

• save -> save_checkpoint

  estimator.save_checkpoint(checkpoint)
  

• load -> load_checkpoint

  estimator.load_checkpoint(checkpoint)
  

3. Summary of the change

Provide new save and load API to support h5 and savemodel format (both load and hdfs).

4. How to test?

• [ ] Unit test
• [x] Test saving/loading on HDFS
• [x] Test on Yarn