Redefine Task to support different AI jobs

[skydoorkai]

The current task definition is based on data-parallel deep learning training and assumes that the training data is from a file or a database table represented by a string.

We need to design a more general task definition to support:

1. Different kinds of AI jobs, such as XGboost.

2. Custom-defined data read.

3. No need to change the proto definition when adding new types of AI jobs in ElasticDL.