Make all implicit re-exports explicit

[e10e3]

Currently, River allows most modules to use implicitly re-exported variables (the error check is disabled).
It turns out there were only two cases where MyPy found this error, both in the optim module. I chose to re-export both because it would make sense to load them from these files (this is in fact where they were originally defined, before a refactor).

With these cases fixed, the rule can now be enabled everywhere.

[e10e3]

This change just showed its use: a previous PR used a tangled import instead of directly importing a module.
I should have rebased before pushing.

Let me fix that.

[e10e3]

There's an error in the doctest of multioutput.chain.ClassifierChain, saying “OpenMLError: Dataset yeast with version 4 not found.” I can't reproduce it on my side. Maybe a network error from GitHub?

[MaxHalford]

Does the error persist?

[e10e3]

Does the error persist?

I still can't reproduce the error on my side, making me think it's a transient error. The only way to know if it's still present in CI is to re-run the job (but I need someone with the rights to do so).