Preserve extension dtype in str.split(expand=True) (fixes #11884) - #12482

What

Series.str.split(..., expand=True) on a string / string[pyarrow] column returns object columns instead of preserving the input dtype, which is inconsistent with pandas.

Fixes #11884.

Why

SplitMap._meta rebuilds a one-row sample so the split produces the right number of columns, but constructs it with the default dtype:

meta = self.frame._meta._constructor(
    [delimiter.join(["a"] * (self.n + 1))],
    index=meta.iloc[:1].index,
)

_constructor([...]) defaults to object, so meta.str.split(expand=True) runs on an object series and the resulting meta — and every partition's output — becomes object.

Fix

Pass the source dtype when building the sample so the split preserves it:

             index=meta.iloc[:1].index,
+            dtype=self.frame._meta.dtype,

Test

Added test_str_split_expand_preserves_dtype: splits a dtype="string" column with expand=True and asserts (via assert_eq, which checks dtypes) that the result matches pandas. It fails on main (object) and passes with this change. Uses the pandas string dtype, so no pyarrow dependency is required.