BUG: DataFrame.rank() loses ExtensionArray dtype (GH#52829)

[maheshmakvana]

Problem

DataFrame.rank() silently converts ExtensionArray-backed columns (e.g. ArrowDtype, nullable Int64) to float64, while Series.rank() correctly preserves the original dtype.

import pandas as pd
import pyarrow as pa

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

s.rank(method="min").dtype        # uint64[pyarrow]  checkmark
df.rank(method="min")["a"].dtype  # float64  before fix
                                  # uint64[pyarrow]  after fix

Root cause

In NDFrame.ranker(), the DataFrame branch used data.values which consolidates all blocks into a single numpy array, stripping EA type info.

Fix

For axis=0 (default), use _mgr.apply() to process each block independently:

• 1-D EAs (ArrowExtensionArray, IntegerArray, etc.) call their own _rank(), preserving dtype.
• numpy blocks and 2-D EAs (DatetimeArray) are transposed, ranked with algos.rank(axis=0), and transposed back.
• axis=1 keeps the old data.values path (cross-column ranking must see all columns at once).

Closes #52829