Fix inference for xgboost fitted with early-stopping

[avolkov-intel]

Description

Earlier there was a difference in results between xgboost predictions and predictions of our modelbuilder if xgboost was fitted using early-stopping parameter. This pull request is to fix this issue

[Alexsandruss]

/intelci: run

[avolkov-intel]

@Alexsandruss can we merge it?

[napetrov]

Also a broader question - can we add tests for this case? we might do this as a second iteration after enabling fix due to release timelines

[avolkov-intel]

/intelci: run

[avolkov-intel]

Also a broader question - can we add tests for this case? we might do this as a second iteration after enabling fix due to release timelines

Yes, I think we should do this in different pull request to merge that one before release

[samir-nasibli]

Also a broader question - can we add tests for this case? we might do this as a second iteration after enabling fix due to release timelines

Yes, I think we should do this in different pull request to merge that one before release

I am expecting that we are testing the fix in the PR, so it is better to add simple test case in this PR as well.

[napetrov]

Manual testing is fine for this PR integration, but should be added anyway

[razdoburdin]

Looks good