Closes gh-350

Otherwise, the version which emulates "clip" fails with torch.vmap.

This patch is less innocuous than it looks, because it changes the promotion rules. Previously, min and max were not upcasting the result:

In [9]: xpn = array_namespace(np.arange(3))

In [10]: xpn.clip(xpn.arange(5, dtype=xpn.int8), 2, xpn.asarray(3.0))
Out[10]: array([2, 2, 2, 3, 3], dtype=int8)

and now they are:

In [11]: xpt = array_namespace(torch.ones(3))

In [12]: xpt.clip(xpt.arange(5, dtype=xpt.int8), 2, xpt.asarray(3.0))
Out[12]: tensor([2., 2., 2., 3., 3.])

OTOH, the wording in the spec, https://data-apis.org/array-api/draft/API_specification/generated/array_api.clip.html is

min ... should have the same data type as x.

and

If either min or max is an array having a different data type than x, behavior is unspecified and thus implementation-dependent.