Warm start to save time when retraining estimators

**Description**
"Warm start" in this case refers to a) training a model on training data, b) pausing to do stuff like generate predictions on validation data and then c) continuing the training process on more training data, using the model trained in the first step as the starting point.

**Use-case**
Suppose we define an automl algorithm which gives the best pipelines iteratively more and more data to train on.

In that case, retraining with "warm start" on the new training data could save time over retraining from scratch on the old + new training data.

Some estimators may not support this easily, so our API should be tolerant to that.