Improve fitting a pipeline to the dataset

[ravinkohli]

The fit method currently does not honour the user constraints set while initialising or running the search. These include:

1. Memory constraint
2. Function time limit
3. Resampling strategy
4. include and exclude components and search space updates.

Also in its current format, it can not train the pipeline using cross validation.

We already have the infrastructure to do so, i.e, the TAE similar to how dummy and traditional predictions are done.

For an idea, we can also use auto-sklearn's implementation for this.