Batch in arrow when the source is arrow-backed, and let _consolidate type numpy scalars

[Yonghui-Lee]

Fixes #8663

1. IterableDataset.batch() takes arrow-backed data out of arrow for no reason
(iterable_dataset.py). It regroups via a Python transpose, _batch_fn, which converts arrow
to Python examples and transposes them back. #8126 already added an arrow path for this, gated
on self._formatting.is_table; this extends the gate to any arrow-backed source:

-        if self._formatting and self._formatting.is_table:
+        if self._formatting and (self._ex_iterable.iter_arrow is not None or self._formatting.is_table):

2. _consolidate cannot type numpy scalars (np_formatter.py):

-                isinstance(x, np.ndarray) and x.shape == column[0].shape and x.dtype == column[0].dtype
+                isinstance(x, (np.ndarray, np.number, np.bool_))
+                and x.shape == column[0].shape
+                and x.dtype == column[0].dtype

Performance

200k rows x 4 columns, in-memory source, single process, pinned with taskset. Absolute wall time for a full pass of .batch(n).

Arrow-backed sources:

format	bs	before (s)	after (s)	speedup
numpy	32	6.663	2.775	2.4x
numpy	256	4.430	0.383	11.6x
numpy	1000	4.147	0.105	39.5x
numpy (string)	1000	3.633	1.306	2.8x
torch	32	12.020	3.128	3.8x
torch	256	8.830	0.423	20.9x
torch	1000	8.577	0.127	67.5x
no format	32 / 256 / 1000	0.961 / 0.622 / 0.579	0.991 / 0.611 / 0.574	unchanged

Generator-backed sources

format	bs	before (s)	after (s)
numpy	32 / 256 / 1000	13.253 / 12.272 / 12.516	13.382 / 12.372 / 12.556
torch	1000	18.820	18.309
no format	32 / 1000	0.335 / 0.280	0.330 / 0.273

[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.