BUG: DataFrame.rank does not preserve ExtensionArray dtypes

[weeknd415]

• closes BUG: DataFrame.rank does not return EA types when original type was an EADtype #52829
• Tests added and passed
• All CI tests passed
• Added entry in doc/source/whatsnew/ (follow the existing format)

Summary

DataFrame.rank() converts PyArrow-backed and nullable ExtensionArray columns to float64, while Series.rank() correctly preserves the EA dtype. This is because the internal ranker() function calls data.values for 2D data (DataFrames), which goes through BlockManager.as_array() and strips all ExtensionArray type information.

Reproducer

import pandas as pd
import pyarrow as pa

s = pd.Series([1, 2, 3], dtype=pd.ArrowDtype(pa.int32()))
df = s.to_frame(name="a")

print(s.rank(method="min").dtype)       # uint64[pyarrow] ✓
print(df.rank(method="min").dtypes)     # float64 ✗ (should be uint64[pyarrow])

Fix

Replace the ranker() closure with block-level processing via self._mgr.apply(), following the same pattern used by _accumulate(). This processes each block independently:

• ExtensionArray blocks → dispatch to EA._rank(), preserving dtype
• NumPy blocks → dispatch to algos.rank(), same as before
• axis=1 (cross-column ranking) → falls back to NumPy conversion since ranking across columns requires a single array

Tests Added

• test_rank_ea_dtype_preservation — PyArrow int32/float64 columns across all 5 rank methods (average, min, max, first, dense)
• test_rank_ea_dtype_preservation_nullable — Nullable Int64/Float64 columns with NA values