Added generalized multiple outputs

[seiyatsukamoto]

Changed the export and infer to map multiple outputs.
Changed Ledger to accept dictionaries.
Changed EventSet to take in extra_params.
Added median pool for regression outputs.
Changes shouldn't require changes for single output configs.
Changes were checked with timedomain, merger time regression, and multitask branches.

[wbenoit26]

I'll give this a look tomorrow. Could you explain the motivation for these changes?

[seiyatsukamoto]

It's for #474.
I also think that saving multiple outputs should be somewhere for the BBH + BNS runs.

[wbenoit26]

I think we can simplify the ledger changes quite a bit by subclassing EventSet rather than adding dictionaries and extra params:

class MergerTimeEventSet(EventSet):
    merger_time: np.ndarray = parameter()

We'll have to generalize the recover function, but that should be doable.

There's a couple of bugs I've noticed in the base branch as well that we should get fixed. Here's what I suggest:

1. I'll make a quick PR fixing those bugs.
2. You rebase on top of that commit with the updated Ledger approach.
3. After merging this in, we make a PR against dev to see what it would take to get things merged. We should do a complete test before merging things in, but it would be good to see the diff.

[wbenoit26]

Just merged #511. Give the subclassing approach a try and rebase on top of it.