Victor/krigingbeliever

[vpicheny]

This PR implements the "kriging believer" / "fantasize greedy" strategy. In short, this strategy 1) greedily optimises an acquisition function, 2) "fantasize" a value at the chosen query points (either by using the GP mean at this qp, or by generating a GP sample), 3) update the posterior of the GP, conditioning on this fantasized data, 4) iterate over 1-3.

We do this by introducing a FantasizedGPRModel wrapper. This model can only do predict, not optimize/update.

[vpicheny]

Introduces two breaking changes:

• ModelStack is renamed TrainableModelStack
• file local_penalisation is renamed greedy_batch

[uri-granta]

Looks good. Just a couple of small comments.

[hstojic]

Introduces two breaking changes:

• ModelStack is renamed TrainableModelStack
• file local_penalisation is renamed greedy_batch

LocalPenalizationAcquisitionFunction has been renamed to LocalPenalization

[hstojic]

Few final minor comments, after that good to go!

To follow up with:

1. Add integration tests
2. Test that it works with async BO