Return numpy scalars from numpy formatting and keep duration dtypes

[Kayvan-Zahiri]

Closes #8500.

Two problems in NumpyFormatter, both falling out of numpy's type hierarchy.

Scalars. np.bool_ and np.datetime64 are not np.number, so _tensorize
did not early-return them and _recursive_tensorize ran them through
__array__ first. Row access handed back 0-d ndarrays, which are unhashable
and fail isinstance(x, np.bool_). Both checks now use np.generic, which
covers every numpy scalar; np.character and np.number are already subclasses
of it, so the two existing branches collapse into it cleanly.

Durations. np.issubdtype(np.dtype("timedelta64[us]"), np.integer) is
True, so the int64 default dtype applied to duration columns. That is why the
same column reported timedelta64 by row and int64 by batch. The default now
skips timedelta64.

Before and after, on {"b": True, "t": datetime, "d": timedelta, "i": 3}:

row    b: ndarray(0-d, unhashable) -> np.bool_
       t: ndarray(0-d, unhashable) -> np.datetime64
batch  d: int64                    -> timedelta64[us]

int64 and float32 defaults for ordinary integer and float columns are
unchanged, and so are strings.

Two tests added, both failing without the change. tests/test_formatting.py,
tests/test_arrow_dataset.py and tests/test_table.py stay at 664 passed with
the same single pre-existing failure (test_dataset_to_iterable_dataset, which
fails identically on a clean checkout here). ruff check and ruff format
clean.

Separate from #8518, which touches table_cast.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[ebarkhordar]

#8500 has three open PRs against src/datasets/formatting/np_formatter.py, so I fetched each head's copy of that file from this repo and diffed them.

#8503 5a97bc12  sha256 7210141c7c3c9f13
#8504 86105122  sha256 0cbf9e401c97e48c
#8519 1335ea15  sha256 99d8e045fc467e27

diff 8504 8519 -> 4 added lines, all comments (the np.generic note
                  and the timedelta64 note). No executable difference.
diff 8503 8504 -> 1 line. 8503 leaves _recursive_tensorize checking
                  (np.ndarray, np.character, np.number).

So the code change here is the same one #8504 makes, and #8503 is that change without the _recursive_tensorize line.

Where this PR does differ is the tests. It also asserts format_column, checks each scalar is not an ndarray, calls hash() on it, and covers a plain int column. #8504's single test asserts the three isinstance checks and the batch duration dtype only.

I have no view on which should land. Flagging it so the same diff does not get reviewed three times.