Make WITH clause in TRAIN statement optional

[shendiaomo]

The default setting of LinearClassifier works well in many cases. Omitting the WITH clause is more friendly for novice users.
For example

SELECT * FROM iris.train WHERE class != 2
TO TRAIN LinearClassifier LABEL class INTO sqlflow_models.temp;

will get a decent result as

{'accuracy': 1.0, 'accuracy_baseline': 0.54285717, 'auc': 1.0, 'auc_precision_recall': 1.0, 'average_loss': 0.23411807, 'label/mean': 0.45714286, 'loss': 0.23411807, 'precision': 1.0, 'prediction/mean': 0.56584316, 'recall': 1.0, 'global_step': 70}

while

SELECT * FROM iris.train WHERE class != 2
TO TRAIN DNNClassifier WITH model.hidden_units=[100, 100] LABEL class INTO sqlflow_models.temp;

can only achieve

{'accuracy': 0.7, 'accuracy_baseline': 0.54285717, 'auc': 1.0, 'auc_precision_recall': 1.0, 'average_loss': 0.56149596, 'label/mean': 0.45714286, 'loss': 0.56149596, 'precision': 1.0, 'prediction/mean': 0.4176486, 'recall': 0.34375, 'global_step': 70}

The loss of LinearClassifier with default learning rate and optimizer is less than half of the DNNClassifier with a similar setting.