Treating Ordinals as purely Categorical may make optimizers weaker than they should be

[eddiebergman]

See these lines when converting form ConfigSpace spaces to Neps space:

neps/neps/search_spaces/search_space.py

Lines 33 to 42 in 8b2d498

	elif isinstance(hyperparameter, CS.CategoricalHyperparameter):
	parameter = CategoricalParameter(
	hyperparameter.choices,
	default=hyperparameter.default_value,
	)
	elif isinstance(hyperparameter, CS.OrdinalHyperparameter):
	parameter = CategoricalParameter(
	hyperparameter.sequence,
	default=hyperparameter.default_value,
	)

This may be fine for small ordinals like ["small", "medium", "large"], just treating them as a categorical, but, for tabular benchmarks this may be an issue.

For example, consider a benchmark which only has tabular entries for hyperparameters x, y, i.e.

x = Ordinal([1, 1.5, 4.5, 16, 32.354, ..., 100])
y = Ordinal(["small", "medium", "large"])

An optimizer which takes into account their order should theoretically outperform one which doesn't, i.e. SMAC.

One hacky solution is to convert it to an integer representation that act as indices so order information is presevered?