[ML] Improve data access pattern computing best splits for classification and regression

[tveasey]

The feature derivatives for boosted tree splits are laid out in increasing feature index order, but when we sample a feature bag we don't ensure it is sorted. This means we are getting random access rather than linear scan over the derivatives when computing the best split. This switches to sorting the feature bag after sampling and rejigs the function signatures to avoid allocating a vector for every node.

[droberts195]

LGTM

[tveasey]

retest

[tveasey]

Thanks for the review @droberts195! While I was in the area I realised:

1. I could avoid all allocations in categorical sample without replacement.
2. There was a bug in the case more samples than values were asked for: missing return. (This wasn't actually being exercised in the code, but it is clearly worth fixing.)

I also added a unit test. I decided it isn't worth breaking this out into a separate change (it is also essentially related to trying to improve cache performance), but it is worth also having a look at e548cc6

[droberts195]

Still LGTM

Neither of the comments is essential to change if the CI goes green and you need to merge quickly