`NpyDatetime_ConvertPyDateTimeToDatetimeStruct` is the per-element worker
behind casting object arrays of Python datetimes to datetime64 (e.g.
`np.asarray(list_of_datetimes).astype("M8[us]")`, the floor of
matplotlib's date2num).  It extracted the seven date/time fields with a
`PyObject_HasAttrString` + `PyObject_GetAttrString` pair per field, each a
generic `__getattribute__` dispatch that builds a temporary key string and
a temporary int.

Two complementary changes:

* Fast path: for genuine datetime.date / datetime.datetime objects
  (including subclasses such as pandas Timestamp), read the fields straight
  from the CPython C struct via the datetime C-API macros, with no
  attribute lookups at all.

* For the remaining attribute accesses (the duck-typed fallback and the
  per-element ``tzinfo`` lookup), use a single ``PyObject_GetOptionalAttr``
  on pre-interned attribute names (added to the ``npy_interned_str``
  registry), fusing the presence check and the lookup and avoiding the
  temporary key object.  The ``tzinfo`` lookup runs for every datetime, so
  this matters even when the fast path supplies the fields.

Both paths converge on a single shared validation tail (``has_time`` flag),
so the date/time range checks and tzinfo handling exist in one place.
Output is bit-identical to before for naive, tz-aware, and date inputs.

Regression tests cover the datetime/date subclass fast paths, the
duck-typed fallback (which real datetimes no longer exercise), invalid
dates, and the tz-aware fast path.

Co-Authored-By: Claude Opus 4.8 (1M context) <[email protected]>