only evaluate items of builtin collections not subclasses

[jeromedockes]

When evaluating a DataOp, we evaluate built-in collections such as lists by evaluating each item and then reconstructing the list. This allows things like choose_from([choose_from(...), choose_from(...)]) (because we recurse into the list) and many more useful cases. This is similar to what scikit-learn clone does.

However, we should not do it for instances of subclasses of those builtin-types, because their initializers may work differently or they may have other invariants we do not know about so we may not be able to reconstruct the new collection from the evaluated items. sklearn.clone only considers the built-in classes, not their subclasses.

At the moment we also open subclasses, this PR fixes that (ie makes the behavior the same as clone)

eg this fails on main but works in this PR

>>> from collections import namedtuple

>>> import skrub

>>> Pair = namedtuple("Pair", ["first", "second"])
>>> p = Pair(1, 2)
>>> skrub.as_data_op(p).skb.eval()
Pair(first=1, second=2)

(In main we get TypeError: Pair.__new__() missing 1 required positional argument: 'second' when trying to call Pair([1, 2]) )

[rcap107]

LGTM, thanks @jeromedockes

I wonder how many people were affected by this, besides you 😝

I noted down this should end up in the documentation somehow, maybe as a possible error