[ML] Improvements to regression and classification memory handling

[tveasey]

Currently, our upfront memory usage estimate is an upper bound and a significant over estimate in most cases. This issue covers a couple of quick wins for improving the estimate:

• Pass the training percentage: we need to know the number of rows used to train
• Pass number of feature values for each feature: this would enable to better estimate how much memory we'll use for aggregate loss derivatives
• Account for maximum number of features we will select
• The SHAP's memory usage is not current with the leaf statistics memory usage

A better strategy (longer term) would be, rather than estimating a memory upper bound, estimate a value which training is very unlikely to exceed. This would require that we support circuit breaking during training. Since we snapshot state periodically the user would still be able to retrospectively increase the memory limit and restart analysis.

[tveasey]

This is also impacted by #1003.

[tveasey]

We've done quite a lot of work on memory usage since this issue was created. Whilst it would be possible to refine estimates if we knew certain features had relatively few distinct values this is not a priority at present. We can revisit if we decide we need better memory estimates in the future.