Preserve torch autograd in the batched cov path

The generic `cov` implementation called `xp.asarray(m)` on its input. For
torch this detaches gradients and mutates the caller's tensor in place, so
`cov` on a batched tensor (ndim > 2, which routes to the generic path) with
`requires_grad=True` returned a detached result and silently zeroed the
input's grad.

The call is unnecessary: the delegation layer already guarantees `m` is an
array (it calls `array_namespace(m)` and reads `m.ndim`). Drop it, and add
a torch autograd regression test.

Author: bruAristimunha

src/array_api_extra/_lib/_funcs.py

@@ -270,7 +270,9 @@ def cov(
     xp: ArrayNamespace,
 ) -> Array:  # numpydoc ignore=PR01,RT01
     """See docstring in array_api_extra._delegation."""
-    m = xp.asarray(m)
+    # NB: no `xp.asarray(m)` here. The delegation layer already guarantees `m`
+    # is an array (it calls `array_namespace(m)` and reads `m.ndim`), and on
+    # torch `xp.asarray` detaches gradients and mutates the caller's tensor.
     dtype = (
         xp.float64 if xp.isdtype(m.dtype, "integral") else xp.result_type(m, xp.float64)
     )

tests/test_funcs.py

@@ -801,6 +801,22 @@ def test_weights_wrong_length(self, xp: ModuleType):
         with pytest.raises((ValueError, RuntimeError)):
             _ = cov(m, correction=0.5, aweights=w_bad)
 
+    def test_torch_autograd(self, torch: ModuleType):
+        # The batched (generic) path must not detach gradients or mutate the
+        # input tensor in place, as `xp.asarray` does on torch.
+        xp = torch
+        rng = np.random.default_rng(20260417)
+        m = xp.asarray(rng.random((4, 3, 20)), dtype=xp.float64)
+        m.requires_grad_(True)
+        # cov returns the array-api `Array` type; at runtime it is a torch
+        # tensor, so cast to access autograd attributes without type errors.
+        c = cast(Any, cov(m))  # batched -> generic path
+        assert c.requires_grad
+        assert m.requires_grad  # input tensor not mutated
+        c.sum().backward()
+        assert m.grad is not None
+        assert bool(xp.all(xp.isfinite(m.grad)))
+
 
 @pytest.mark.xfail_xp_backend(Backend.SPARSE, reason="no arange", strict=False)
 class TestOneHot: