Support Keras functional API in SQLFlow

[typhoonzero]

This is also a part of sql-machine-learning/models#22, to unify custom model for both SQLFlow and ElasticDL.

Need to update the document https://github.com/sql-machine-learning/sqlflow/blob/develop/doc/design/design_customized_model.md if the feature is implemented.

To use the feature column together with Keras functional API please refer to tensorflow/tensorflow#27416 (comment) for implementation detail.

[typhoonzero]

PS: we can implement custom model functional API like https://github.com/sql-machine-learning/elasticdl/blob/develop/model_zoo/cifar10_functional_api/cifar10_functional_api.py. The function returns tf.keras.Model should have a more meaningful name describing the current model, like MyDNNClassifier etc.