[Optimizer] Multi-fidelity algorithms should explicitly check for a fidelity parameter

[eddiebergman]

Right now if running SH with no fidelity parameter available, it will raise this error

        super().__init__(
            pipeline_space=pipeline_space,
            budget=budget,
            loss_value_on_error=loss_value_on_error,
            cost_value_on_error=cost_value_on_error,
            ignore_errors=ignore_errors,
            logger=logger,
        )
        if random_interleave_prob < 0 or random_interleave_prob > 1:
            raise ValueError("random_interleave_prob should be in [0.0, 1.0]")
        self.random_interleave_prob = random_interleave_prob
        self.sample_default_first = sample_default_first
        self.sample_default_at_target = sample_default_at_target
    
>       self.min_budget = self.pipeline_space.fidelity.lower
E       AttributeError: 'NoneType' object has no attribute 'lower'

This is not very descriptive on how to solve the problem. If implementing #119, we can simply check this and emit it in the BaseOptimizer.

Also seems to occur from hyperband_custom_default while not happening for hyperband. Priorband requires it too.