zonal.stats: count returns 0 for empty zones (other stats stay NaN)

[brendancol]

Closes #2644

What this does

zonal.stats() used to report NaN for every statistic of an "empty" zone (one
that exists in the zones raster but has no valid values after filtering NaN and
nodata_values), including count. This makes count return 0 for empty
zones instead, while every other statistic stays NaN.

• count is a cardinality: zero valid cells means a count of 0, not undefined.
• mean, min, max, sum, std, var, majority, and custom callables
  remain NaN for empty zones, since those are undefined over no values.
• The stats() docstring now documents the empty-zone rule explicitly.

This is a deliberate behavior change. Three tests pinned the old NaN-count
behavior; they now expect 0, with comments and a commit message explaining
why. Issue #2644 frames the tradeoff (keep NaN vs return 0) and recommends 0
for count specifically.

Backend coverage

numpy, cupy, dask+numpy, dask+cupy. The numpy path passes an empty_zone_value
into _calc_stats; the cupy path handles its size == 0 branch; the dask path
uses a plain nansum count reducer so an all-empty zone totals 0. crosstab
and apply are untouched (they do not share the affected count code).

Test plan

• test_stats_all_nan_zone (all 4 backends): empty-zone count is 0
• test_stats_all_nan_zone_preserved (numpy/cupy): count 0 for all-NaN zone
• test_stats_nodata_wipes_zone (all 4 backends): count 0 for all-nodata zone
• full test_zonal.py suite passes (169 passed locally; cupy/dask variants
  skip without CUDA)

Skipped steps

No user-guide notebook and no README feature-matrix row: this refines the
documented behavior of an existing function and adds no new public API or
backend support.

[brendancol]

PR Review: zonal.stats count returns 0 for empty zones

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

None. The change is small and the three backend paths line up.

Nits (optional improvements)

• xrspatial/zonal.py _empty_zone_value keys on the literal stat name
  'count'. A custom stats_funcs dict whose key happens to be 'count' but
  whose callable is not the cardinality counter would also get 0 for empty
  zones. That is an unlikely corner and the current behavior is defensible
  (a column named count should behave like a count), so this is just worth a
  mention, not a change.

What looks good

• The empty-zone rule is consistent across numpy (_calc_stats empty_zone_value),
  cupy (the size == 0 branch), and dask (the dedicated _count_reduce).
  Verified locally: numpy and dask both return count 0 and mean/sum NaN for an
  all-NaN zone, including with ragged chunks that split a zone across blocks.
• Variance is unaffected: the dask var merge reads the raw per-block count
  stack, not the reduced count, so changing the reduced count to 0 does not
  perturb std/var.
• The dask mean stays NaN for an empty zone because sum is NaN and NaN/0 is NaN.
• The docstring documents the empty-zone semantics precisely, and the three
  tests that pinned NaN counts were updated to 0 with comments explaining the
  deliberate change.

Checklist

• Algorithm matches intent (count is a cardinality; 0 for empty)
• All implemented backends produce consistent results
• NaN handling is correct (other stats stay NaN)
• Edge cases covered by tests (all-NaN zone, all-nodata zone)
• Dask chunk boundaries handled correctly (verified with ragged chunks)
• No premature materialization or unnecessary copies
• Benchmark not needed (no new function, no hot-path change)
• README feature matrix not applicable (no new function)
• Docstring present and accurate