Suggestion: Allow overriding default training parameters in AutoML

**Is your feature request related to a problem? Please describe.**
Right now autoML uses defaults which perform generally well. I'd like to be able to run a faster "bootstrap run" for big datasets.
There already exists mlContext.Data.BootstrapSample for the data.

However, I am not able to re-run AutoML with full data starting from here bootstrap sample data left off. Running AutoML with full data begins from the defaults again.

There are also other cases where allowing to set default (preferably a list of defaults) might be useful. For example, I might already have run AutoML for the previous version of the dataset. The new dataset might have some columns added and removed, so I'd like to start with those defaults.

**Describe the solution you'd like**
AutoML should accept a parameter for defaults to start training from. Ideally, I would be able to submit multiple parameters for each trainer, but also just 1 set of training parameters is enough.

**Describe alternatives you've considered**
None. Are there other ways to implement this, except customize Microsoft.ML source, or re-write own implementation of AutoML?

Maybe it should be possible to also submit an incomplete set of training parameters, in case developer only wants to affect one of them, such as reduce learning rate.