Fix NixIO writing of datetime array_annotations (#1198)

[Leonard013]

Fixes #1198.

Why

NixIO cannot write array_annotations whose elements are datetime (or
date/time) objects. neo.core accepts such values as array-annotation
elements (they are in ALLOWED_ANNOTATION_TYPES, stored as an object-dtype
ndarray), so a Segment/Event/AnalogSignal carrying them can be built and
lives fine in memory — but writing it to a .nix file crashes:

File ".../neo/io/nixio.py", line 1283, in _write_property
    section.create_property(name, values)
File ".../nixio/section.py", line 156, in create_property
    dtype = DataType.get_dtype(single_val)
File ".../nixio/datatype.py", line 50, in get_dtype
    raise ValueError("Unknown type for value {}".format(value))
ValueError: Unknown type for value 2021-01-01 00:00:00

As @JuliaSprenger pointed out in the issue, _write_property must convert the
datetime objects before saving. Neo already knows how to do this for scalar
datetime annotations via dt_to_nix() / dt_from_nix(); the iterable
(array-annotation) branch just never used them.

What

neo/io/nixio.py, two surgical changes:

1. Write (_write_property, iterable branch): when an item is a
   datetime/date/time, convert it to its ISO string with dt_to_nix()
   and record the returned annotation type as the property's definition —
   exactly mirroring the existing scalar-annotation path a few lines above.
2. Read (_nix_attr_to_neo): when a property carries a datetime
   definition and holds multiple values (a list), reconstruct it
   element-wise with dt_from_nix() into an object-dtype ndarray. A single
   value keeps the previous scalar behaviour unchanged. The object dtype is
   required so the reconstructed annotation is accepted back by
   array_annotate() on the read path.

This makes datetime array_annotations survive a write → read roundtrip for all
three of datetime, date and time.

Tests

Added NixIOWriteTest.test_datetime_array_annotations in
neo/test/iotest/test_nixio.py (modelled on the neighbouring
test_empty_array_annotations): writes an Event with datetime, date and
time array_annotations, reads it back, and asserts equality. The test
fails on master (same ValueError: Unknown type for value ...) and
passes with this change.

Verification (Python 3.12, nixio 1.5.4, numpy 2.5.1):

black --check (changed files)   -> 2 files left unchanged
pytest new regression test      -> 1 passed
pytest full test_nixio.py       -> 36 passed, 7 skipped  (skips need datalad web downloads)
pytest core dataobject          -> 5 passed
source-only revert of the fix   -> the new test fails with the ValueError above

Scope / caveats for reviewers

• numpy datetime64 (the literal issue title) is intentionally out of
  scope. neo core rejects numpy.datetime64 before it can reach any IO
  (ALLOWED_ANNOTATION_TYPES in neo/core/baseneo.py does not include
  np.datetime64), so the values that actually reach _write_property are
  python datetime objects held in an object-dtype array — which is what the
  issue's own diagnosis ("convert datetime objects before saving") refers to.
  Broadening core validation to accept datetime64 would be a separate, larger
  change.
• The read path distinguishes "array" vs "scalar" datetime purely by
  isinstance(values, list). A length-1 datetime array_annotation is stored and
  read back as a scalar, then normalised to a 1-element object array by
  array_annotate — verified to roundtrip correctly.
• Not touched: the nixrawio/lazy read path — array_annotations roundtrip is a
  full-NixIO feature.

---

Prepared with AI assistance (Claude Code): the reproduction, fix, and
regression test were AI-drafted, and the reasoning and verification are
documented in full above. Happy to adjust the approach or scope.