0.32.0

• perf(array_glyph): count domain cells without materialising a per-cell index list (#305)

• Replace len(get_indices2(frame, [np.nan])), which built one Python tuple
  per cell, with a pure-numpy, mask-aware reduction extracted into
  ArrayGlyph._count_domain_cells.

• • fixes the MemoryError when building a 4-D (n, h, w, 3) RGB animation
    stack: the old len(shape) == 3 frame pick passed the whole stack to
    get_indices2 (~15 GB tuple list); the reduction is O(1) in Python
    objects and ~3000x faster on large frames

• count a stack on frame 0 and a single frame (2-D, or an (h, w, 3) RGB
  image from rgb_bands) whole, using self.rgb to disambiguate the two
  3-D shapes and fix a lone RGB image counting only its first row

• keep the mask term so exclude_value-masked cells stay excluded,
  byte-equivalent to the old get_indices2 (a plain ~np.isnan would
  over-count masked cells)

• add regression tests (4-D, 3-D multi-frame, zero-domain, masked 2-D and
  masked stack, single RGB image, integer dtype) and clarify the
  num_domain_cells docstring

• Closes #304

• fix(styling): drop pandas & numpy null sentinels in categorize (#303)

• categorize's null filter recognised only None and np.nan, so pandas'
  pd.NA / pd.NaT (and, on some numpy builds, datetime64('NaT')) survived
  and became their own colour category -- making categorisation depend on
  the column dtype. Treat every null flavour uniformly so the category set
  depends only on the distinct real values (the in-glyph categorical path
  delegates to categorize and inherits the fix).

  • Drop pd.NA / pd.NaT via pandas' scalar pd.isna, and datetime64('NaT')
    via np.isnan or np.isnat.
  • Keep pandas an undeclared soft dependency: import pd.isna lazily and
    once (not per element), degrading to the numpy-only path (np.isnat for
    datetime NaT) when pandas is absent -- no new runtime dependency.
  • Short-circuit str/bytes and guard np.isnan / pd.isna against array-like
    object-array elements (e.g. range) so odd contents are carried through
    as non-null rather than crashing, matching base behaviour.
  • Extract the check to a module-level _categorical_is_null helper to keep
    categorize's cognitive complexity within bounds.
  • Add tests (pd.NA/pd.NaT, no-pandas fallback, all-null error, mixed null
    kinds, numpy datetime NaT incl. the forced np.isnat path, genuine
    pandas nullable dtypes, array-like element) and refresh the categorize
    / _is_null docstrings.

  Closes #302