FIX keep inherited from builtin containers' types

[adrinjalali]

This turned up in scikit-learn, where a Bunch object would get downgraded to a dict, when the dask backend for joblib was used. We fixed the issue in scikit-learn by not using the Bunch specific features, but the core issue comes from here, and this PR fixes the issue.

• Workaround in scikit-learn: FIX metaedata routing with dask backend on joblib scikit-learn/scikit-learn#34077
• Proposed workaround in joblib: FIX Preserve collection subclasses across the dask backend's implicit scatter joblib/joblib#1798

I think it would make sense to fix the core issue here.

• Tests added / passed
• Passes pixi run lint

fyi @tomMoral @StefanieSenger

---

Claude generated summary, feel free to ignore.

Preserve collection subclasses across Client.scatter

Client.scatter decided whether to treat its argument as a single value or a
collection of items to scatter using isinstance. Because
isinstance(some_dict_subclass, dict) is True, a dict subclass (e.g.
scikit-learn's Bunch) was unpacked as a mapping and returned as {key: Future},
silently losing its type — list/set/tuple subclasses were reconstructed via
input_type(...), but dict subclasses were not.

This switches the collection-vs-single decision to exact-type checks, so any
subclass of a builtin collection is scattered as a single opaque value (pickled,
exact type preserved). Exact dict/list/set/tuple/frozenset keep their
existing "collection of items" behavior.

Background

scikit-learn's metadata routing passes a Bunch (a dict subclass with
attribute access) between estimators. Under joblib's dask backend, arguments
larger than ~1 KB are implicitly scattered, so the Bunch round-tripped through
scatter and arrived on the worker as a plain dict, breaking attribute access
(AttributeError: 'dict' object has no attribute ...).

This was first worked around on the joblib side
(joblib/joblib#1798) by wrapping
such objects before scattering. Review feedback there asked, reasonably, why this
isn't fixed in dask directly — the downgrade happens because scatter chooses to
unpack the object, so scatter should defend its own contract. This PR is that
fix; the joblib-side workaround can be dropped once it lands.

Non-regression for
scikit-learn/scikit-learn#34005.

Tests

• test_scatter_collection_subclass: a dict subclass, list/set subclasses, and a
  namedtuple each scatter to a single Future with their exact type preserved on
  the worker (including attribute access).
• Existing scatter/gather behavior for exact builtin collections is unchanged.

[github-actions]

Unit Test Results

See test report for an extended history of previous test failures. This is useful for diagnosing flaky tests.

    40 files  ± 0      40 suites  ±0   14h 21m 27s ⏱️ + 6m 0s
 4 154 tests + 1   3 973 ✅  -  2    178 💤 ±0  3 ❌ +3 
80 839 runs  +19  76 601 ✅ +18  4 235 💤  - 2  3 ❌ +3 

For more details on these failures, see this check.

Results for commit eb2f68a. ± Comparison against base commit db3cb43.

♻️ This comment has been updated with latest results.

[adrinjalali]

It doesn't look like the failures are related to the PR.