feat(HEBO): Added initial support for single/mutli objective HEBO

[eddiebergman]

NOTE: I will likely merge this soon given I have not heard back from HEBO developers.

DRAFT

This PR implements what I could gather from the HEBO documentation.

So far it seems I could get single objective optimization in and also multi-objective.

There's a few possible issues marked with #TODO, such as HEBO not supporting an unbounded optimal (even if only theoretical) due to it's observe removing non-finite entries. Also tests still need to be added for HEBO space parsing, however it seems to at least function with a basic {"x": (1.0, 10.0)} space.

---

Some other points:

• Should leave an issue to explicitly state their optimizers only minimize in the examples. This wasn't immediately clear.
• Not sure how to really use ref_point in GeneralBO, it seems I can leave it as None and that would seem safer than specifying my own. However if I have the information, it might make GeneralBO more effective to provide it. Not sure how it acts with inf boundaries as the reference point though.

[vlad-moroshan]

May I ask if the "..." operator allows execution to simply fall through to the method defined below?

[eddiebergman]

The ... is essentially just a value, just like writing the number 4. For example,
These are both perfectly valid python code:

def f(x: int):
    4

def f(x: int):
    z = ...

What's making it fall through is the @overload. You can basically think of them as just doc strings that type checkers and language servers can use for better "intellisense". In the example you pointed out, if you pass in a Sequence[Metric], you can expect worst_possible_cost to give you back a list[float]. If you pass in a Metric, you'll get back a float.

Without them, your editor/language server will say, x = worst_possible_cost(single_metric) is either a float or list[float], even though you know it's a float.

[vlad-moroshan]

Thank you for the clarification.