classify: align natural_breaks parameter order with sibling classifiers

[brendancol]

What

natural_breaks ordered its parameters differently from the other
classifiers that take the same arguments. This lines it up and adds the
type hints binary was missing.

Closes #3398.

The drift

quantile(agg, k=4, num_sample=20000, name='quantile')
maximum_breaks(agg, k=5, num_sample=20000, name='maximum_breaks')
natural_breaks(agg, num_sample=20000, name='natural_breaks', k=5)   # before

quantile and maximum_breaks put k first and name last.
natural_breaks did the opposite, so natural_breaks(raster, 5) set
num_sample=5 instead of k=5.

Change

• Reorder natural_breaks to (agg, k=5, num_sample=20000, name=...).
• Add a shim for the old order. Pre-1.0 callers always passed k as a
  keyword (it was the last parameter), so a call that gives k= as a
  keyword and a second positional is using the old order: that
  positional is the old num_sample. Remap it and emit a
  DeprecationWarning. natural_breaks(raster, 20000, k=4) (used in
  one example notebook) keeps working.
• Add agg: xr.DataArray, name: Optional[str], and the return hint to
  binary, the only public classifier without type hints.

Deprecation impact

The reorder is breaking for code that passed num_sample positionally.
The shim covers the one documented pattern (k keyword plus positional
num_sample) and warns rather than silently reinterpreting the value.

Tests

• test_natural_breaks_positional_k_matches_siblings: second positional
  arg is now k.
• test_natural_breaks_legacy_positional_num_sample_warns: the old
  order warns and produces the same result.

Verified on numpy and cupy entry points (CUDA available on this host),
plus the Dataset path. Full test_classify.py and test_validation.py
pass.

Found by the api-consistency deep-sweep against classify.