np.ma.median is inefficient with DataFrames

From:
https://github.com/scikit-learn/scikit-learn/issues/19926

Using np.ma.median in a pandas DataFrame takes a lot longer than np.nanmedian. The issue seems to be about using a full sort rather than a partial argsort.

Example:

import numpy as np, pandas as pd

rng = np.random.default_rng(seed=1)
nrows = int(1e6)
ncols = 10
X = pd.DataFrame(rng.normal(size=(nrows, ncols)))

%%timeit
np.nanmedian(X, axis=0)

191 ms ± 3.98 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

%%timeit
np.ma.median(X, axis=0)

949 ms ± 4.99 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

The problem does not arise when the input is a numpy array:

X2 = rng.normal(size=(nrows, ncols))

%%timeit
np.nanmedian(X2, axis=0)

188 ms ± 2.62 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

%%timeit
np.ma.median(X2, axis=0)

244 ms ± 6.83 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

Start by reproducing the reported benchmarks for np.ma.median and np.nanmedian on a pandas DataFrame and a NumPy array. Investigate the np.ma.median path and verify whether its sorting behavior accounts for the DataFrame-specific slowdown. Done means reducing the DataFrame performance gap while preserving median results.