with_format("numpy") returns bool/datetime values as 0-d unhashable arrays and casts duration columns to int64

[2sumtech]

Describe the bug

With numpy formatting, row access returns np.int64/np.float64/np.str_ scalars for most dtypes, but bool and timestamp/date values come back as 0-dimensional ndarrays (array(True)), which are unhashable and fail isinstance(x, np.bool_) checks. Separately, batch/column access silently casts duration columns from timedelta64[us] to plain int64 (row access keeps np.timedelta64), so the same column reports different types depending on how it's read.

Root cause: NumpyFormatter._tensorize (src/datasets/formatting/np_formatter.py) early-returns only np.number scalars — np.bool_/np.datetime64 aren't np.number, so they fall through to np.asarray() (and _recursive_tensorize re-wraps them via __array__ first). The int64 default-dtype branch uses np.issubdtype(dtype, np.integer), which is True for timedelta64 in numpy's type hierarchy, so duration arrays get the int64 default.

Steps to reproduce the bug

from datasets import Dataset
import datetime as dt
ds = Dataset.from_dict({'b': [True], 't': [dt.datetime(2024,1,1)],
                        'td': [dt.timedelta(seconds=5)]}).with_format('numpy')
row = ds[0]
# actual: b -> array(True) (ndarray!), t -> array('2024-01-01...') (ndarray!)
#         td -> np.timedelta64(5000000,'us') (scalar, inconsistent with batch below)
hash(row['b'])   # TypeError: unhashable type: 'numpy.ndarray'
print(ds[:]['td'].dtype)  # int64  <- duration unit silently stripped

Expected behavior

np.True_, np.datetime64(...) scalars for row access (consistent with int/float/str), and ds[:]['td'].dtype == timedelta64[us] (consistent between row and batch access).

I have a fix ready (7-line change + regression tests) and will open a PR.

Environment info

datasets main (5.0.2.dev0), pyarrow 25.0.1, numpy 2.4.6, Python 3.11.14, macOS