ENH: reshape: test copy= kwarg

[ev-br]

Apparently, copy= kwarg of reshape is not very well reasonably well supported:

• torch does not have it (https://docs.pytorch.org/docs/stable/generated/torch.reshape.html), and array-api-compat simply raises -- fixed by ENH: torch.reshape: implement copy with Torch view and clone operations array-api-compat#457
• cupy will have it soon-ish (in cupy==14.0 via [backport] ENH: add explicit copy arg to reshape cupy/cupy#10147); for now, there's a workaround in array_api_compat
• jax.numpy has it but always copies anyway (unless jit elides the copy)

Where the argument is supported, the test itself needs work:

• on -strict, zero-sized arrays fail some internal check with copy=True -- fixed by BUG: reshape accept an empty array and copy=False array-api-strict#232
• on numpy, the test generates arrays or numpy scalars, and scalars choke something in the test itself. -- fixed by ENH: use native numpy.reshape for numpy>=2.1 array-api-compat#454

[ev-br]

cross-ref #108

[ev-br]

Cupy issue to add the copy= argument to reshape: cupy/cupy#10039

[ev-br]

The array-api-strict failure seems to be

(Pdb) l
118  	
119  	    data = x._array
120  	    if copy:
121  	        data = np.copy(data)
122  	
123  ->	    reshaped = np.reshape(data, shape)
124  	
125  	    if copy is False and not np.shares_memory(data, reshaped):
126  	        raise AttributeError("Incompatible shape for in-place modification.")
127  	
128  	    return Array._new(reshaped, device=x.device)
(Pdb) p data
array([], dtype=float64)
(Pdb) p np.shares_memory(data, data)
False

which a kind of makes sense: if there's no data to share, nothing's shared. So it'll need a tweak in array-api-strict.

[ev-br]

It looks like that after data-apis/array-api-strict#232 and data-apis/array-api-compat#454 by @stanbot8, the only remaining blocker is torch.

[ev-br]

All wrapped modules have been fixed (thanks @stanbot8!), the updated test passes locally. We'll be able to improve further when

• numpy 2.1 is the minimum supported version (will be able to drop branching on the numpy version and always use the native numpy function)
• cupy 14 is the minimum supported version (will be able to drop the wrapper and use cupy's native copy=).