rolling forecast

Allow a time series forecasting model to predict beyond the horizon after new data arrive, without refitting the model. For example,
```python
# predict for the first time, assuming the time horizon = 1 hour. The last timestamp in the training data is 7am.
y_pred_8am = automl.predict(X_pred_8am) # X_pred_8am's timestamp = 8am
# after y_true_8am is observed
X_roll, y_roll = X_pred_8am, y_pred_8am
y_pred_9am = automl.predict(X_pred_9am, X_roll, y_roll)
# after y_true_9am is observed
X_roll = concat(X_roll, X_pred_9am)
y_roll = concat(y_roll, y_pred_9am)
y_pred_10am = automl.predict(X_pred_10am, X_roll, y_roll)
...
```

@int-chaos @egor-bryzgalov @markharley @slhuang @qingyun-wu @luigif2000 @MichalChromcak Does this feature make sense?