Document where the filters differ from scipy.ndimage.

[roytsmart]

Two documentation changes; no behavior change.

Correct the note on even-sized median footprints

The note added in #34 had the comparison backwards. For a footprint with an
even number of elements, scipy.ndimage.median_filter() returns the
larger of the two middle elements, not the smaller, because it delegates
to _rank_filter() and asks for the element of rank size // 2:

a = np.array([0., 10., 20., 30., 40.])
scipy.ndimage.median_filter(a, size=2, mode="nearest")
# array([ 0., 10., 20., 30., 40.])   <- the upper element of each pair

where averaging each pair, as numpy.median() and this package do, gives
[0., 5., 15., 25., 35.]. Selecting rather than averaging is what lets
SciPy return a result in the dtype of the input.

Add a "Differences from scipy.ndimage" section

The README and the documentation landing page both open by saying the
filters are similar to those in scipy.ndimage, and then list only the ways
this library does more. That sets up a migrating user to be surprised, and a
Notes block on one function is only read by someone who already suspects a
difference. The new section covers the other half:

• Boundary modes. "reflect", "constant", and the "grid-*" modes
  raise a ValueError here; "truncate" has no SciPy equivalent.
• Integer input. Promoted to floating point rather than returned in the
  input dtype, which is what allows a fully-excluded where footprint to
  return NaN (Fix the "mirror" boundary mode and validate mode and size. #32).
• Even-sized median footprints. The tie-breaking difference above,
  including that SciPy's convention is a biased estimator: on unit-variance
  noise a size=2 median filter shifts the signal by roughly 0.57, while
  the two conventions agree exactly for odd-sized footprints.

size=2  input mean +0.00013   scipy +0.56635   ndfilters +0.00013
size=3  input mean +0.00018   scipy +0.00032   ndfilters +0.00032
size=4  input mean +0.00423   scipy +0.30070   ndfilters +0.00399
size=5  input mean -0.00042   scipy +0.00031   ndfilters +0.00031
size=8  input mean +0.00087   scipy +0.15258   ndfilters +0.00032

The section states the tradeoff in both directions rather than only ours:
SciPy's rank selection never introduces a value that was not already in the
footprint, which averaging does (a 4-level image filtered at size=4 comes
back with 7 levels).

🤖 Generated with Claude Code

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 100.00%. Comparing base (d71d84c) to head (2c570b1).

Additional details and impacted files

@@            Coverage Diff            @@
##              main       #36   +/-   ##
=========================================
  Coverage   100.00%   100.00%           
=========================================
  Files           14        14           
  Lines          677       677           
=========================================
  Hits           677       677           

Flag	Coverage Δ
unittests	100.00% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.