ENH: fix is_jax_array under torch.compile

[lucascolley]

This fixes errors like the following:

E               torch._dynamo.exc.UserError:
GetAttrVariable(TupleVariable(length=0), dtype) has no type
E               For more information about this error, see:
https://pytorch.org/docs/main/generated/exportdb/index.html#unknown-python-type
E
E               from user code:
E                  File
"/Users/lucascolley/ghq/github.com/data-apis/array-api-extra/src/array_api_extra/_lib/_helpers.py",
line 612, in inner
E                   res = func(*args, **kwargs)  # pyright:
ignore[reportCallIssue]
E                 File
"/Users/lucascolley/ghq/github.com/data-apis/array-api-extra/tests/main/test_at.py",
line 47, in at_op
E                   return meth(y, copy=copy, xp=xp)
E                 File
"/Users/lucascolley/ghq/github.com/data-apis/array-api-extra/src/array_api_extra/_at.py",
line 374, in add
E                   return self._op(_AtOp.ADD, operator.iadd,
operator.add, y, copy=copy, xp=xp)  # pyright:
ignore[reportUnknownArgumentType]
E                 File
"/Users/lucascolley/ghq/github.com/data-apis/array-api-extra/src/array_api_extra/_at.py",
line 289, in _op
E                   (_compat.is_dask_array(idx) or
_compat.is_jax_array(idx))
E                 File
"/Users/lucascolley/ghq/github.com/data-apis/array-api-extra/.pixi/envs/tests-backends/lib/python3.14/site-packages/array_api_compat/common/_helpers.py",
line 253, in is_jax_array
E                   or _is_jax_zero_gradient_array(x)
E                 File
"/Users/lucascolley/ghq/github.com/data-apis/array-api-extra/.pixi/envs/tests-backends/lib/python3.14/site-packages/array_api_compat/common/_helpers.py",
line 84, in _is_jax_zero_gradient_array
E                   cls = cast(Hashable, type(dtype))

cc @crusaderky for vis

[ev-br]

Is there an MWE, ideally as a test or as an example for the added comment?

attrs on empty tuples is a tad cryptic, and I'm not sure how to trigger it---or check if the whole _is_zero_grad_array can be removed, as the comment above indicates will be possible "at some point".

[lucascolley]

calling is_jax_array on a torch array while running under torch.compile(fullgraph=True) should reproduce.

[ev-br]

Great. Add an otherwise failing code snippet to the comment?