Add support for storing the optimization targets and direction of an experiment

[bpkroth]

This PR is useful for mlos-viz and dabl wrapper (#624) to be able to automatically graph the results for a given optimization target, for instance via something like the following:

for opt_target in exp.objectives:
    dabl.plot(exp.results, opt_target)

Since the prior efforts on capturing this data in the Trial metadata are somewhat problematic (allow conflicting changes between runs of an experiment, don't support multi-objective), we extend them to also store values directly as a part of the Experiment, which is a somewhat more appropriate location. Upon retrieval, an attempt is also made to merge the two data sources for backwards compatibility.

This PR does not enforce strictness on that metadata, but future versions could (e.g., disallow resuming an Experiment if it looks like the objective targets have changed. In that case the prior Trial results can potentially still be used to prewarm a new Experiment's optimizer).

[motus]

wait, we already store that data in the trial_param table - see

MLOS/mlos_bench/mlos_bench/run.py

Lines 120 to 124 in 2032d79

	trial = exp.new_trial(tunables, config={
	"optimizer": opt.name,
	"opt_target": opt.target,
	"opt_direction": "min" if opt.is_min else "max",
	})

you can add more key/value pairs to store with each trial there

[bpkroth]

wait, we already store that data in the trial_param table - see

MLOS/mlos_bench/mlos_bench/run.py

Lines 120 to 124 in 2032d79

	trial = exp.new_trial(tunables, config={
	"optimizer": opt.name,
	"opt_target": opt.target,
	"opt_direction": "min" if opt.is_min else "max",
	})

you can add more key/value pairs to store with each trial there

We had a conversation about this offline. I think the conclusion was that:

1. The optimization configuration is really a property of the Experiment and not the Trial, so if something about the Experiment's optimization goal's configuration changes, we should really require a new Experiment, though previously we didn't store or enforce this.
2. That doesn't preclude reusing some TrialData from a previous Experiment to pre-warm a new one.
3. However, storing metadata about the Optimization goals used during a given Trial, which potentially redundant, may still be useful.
   Though, with that in mind, we may also want to store other parameter data used too ...

I think with that in mind, I will still pursue this PR and a future one can enforce the strictness aspects of not changing things, perhaps after we make it easier to merge other "related" TrialData (though capturing which those are might itself be tricky).

[motus]

Looks good! I leave it up to you to keep the optimization direction nullable or roll back to the non-nullable variant