BUG: limit the range of elements in test_{eig,eigvals}

Otherwise torch fails to converge if the matrix is too ill-conditioned.
Notice 6.7e+111, 6.9e+128, 7e+90 values in this failing example:

```
>       res = linalg.eigvals(x)
E       torch._C._LinAlgError: torch.linalg.eigvals: (Batch element 1): The algorithm failed to converge because the input matrix is ill-conditioned or has too many repeated eigenvalues (error code: 6).
E       Falsifying example: test_eigvals(
E           x=tensor([[[ 0.0000e+00,  2.0000e+00,  2.0000e+00,  2.0000e+00,  2.0000e+00,
E                      2.0000e+00],
E                    [ 2.0000e+00,  2.0000e+00,  2.0000e+00,  2.0000e+00,  2.0000e+00,
E                      2.0000e+00],
E                    [ 2.0000e+00,  2.0000e+00,  2.0000e+00,  2.0000e+00,  2.0000e+00,
E                      2.0000e+00],
E                    [ 2.0000e+00,  2.0000e+00,  2.0000e+00,  2.0000e+00,  2.0000e+00,
E                      2.0000e+00],
E                    [ 2.0000e+00,  2.0000e+00,  2.0000e+00,  2.0000e+00,  2.0000e+00,
E                      2.0000e+00],
E                    [ 2.0000e+00,  2.0000e+00,  2.0000e+00,  2.0000e+00,  2.0000e+00,
E                      2.0000e+00]],
E
E                   [[ 2.0000e+00,  2.0000e+00,  2.0000e+00,  2.0000e+00,  2.0000e+00,
E                      2.0000e+00],
E                    [6.7525e+111,  2.0000e+00,  2.0000e+00, 6.9317e+128,  2.0000e+00,
E                      2.0000e+00],
E                    [ 2.0000e+00,  2.0000e+00,  2.0000e+00,  2.0000e+00,  2.0000e+00,
E                      2.0000e+00],
E                    [ 2.0000e+00,  0.0000e+00,  2.0000e+00,  2.0000e+00,  7.5592e+90,
E                      2.0000e+00],
E                    [ 2.0000e+00,  2.0000e+00,  2.0000e+00,  2.0000e+00,  2.0000e+00,
E                      2.0000e+00],
E                    [ 2.0000e+00,  2.0000e+00,  2.0000e+00,  2.0000e+00,  2.0000e+00,
E                      2.0000e+00]]], dtype=torch.float64),
E       )

array_api_tests/test_linalg.py:375: _LinAlgError
```

Author: ev-br

array_api_tests/hypothesis_helpers.py

@@ -331,11 +331,14 @@ def matrix_shapes(draw, stack_shapes=shapes()):
 square_matrix_shapes = matrix_shapes().filter(lambda shape: shape[-1] == shape[-2])
 
 @composite
-def finite_matrices(draw, shape=matrix_shapes()):
-    return draw(arrays(dtype=floating_dtypes,
-                           shape=shape,
-                           elements=dict(allow_nan=False,
-                                         allow_infinity=False)))
+def finite_matrices(draw, shape=matrix_shapes(), dtype=floating_dtypes, bound=None):
+    # for now, only generate elements from (1, bound); cf symmetric_matrices
+    elements = dict(allow_nan=False, allow_infinity=False)
+    if bound is not None:
+        elements.update(**dict(min_value=1, max_value=bound))
+
+    return draw(arrays(dtype=dtype, shape=shape, elements=elements))
+
 
 rtol_shared_matrix_shapes = shared(matrix_shapes())
 # Should we set a max_value here?

array_api_tests/test_linalg.py

@@ -335,7 +335,7 @@ def test_eigvalsh(x):
 @pytest.mark.unvectorized
 @pytest.mark.xp_extension('linalg')
 @pytest.mark.min_version("2025.12")
-@given(x=arrays(dtype=all_floating_dtypes(), shape=square_matrix_shapes))
+@given(x=finite_matrices(dtype=all_floating_dtypes(), shape=square_matrix_shapes, bound=10))
 def test_eig(x):
     res = linalg.eig(x)
 
@@ -370,7 +370,7 @@ def test_eig(x):
 @pytest.mark.unvectorized
 @pytest.mark.xp_extension('linalg')
 @pytest.mark.min_version("2025.12")
-@given(x=arrays(dtype=all_floating_dtypes(), shape=square_matrix_shapes))
+@given(x=finite_matrices(dtype=all_floating_dtypes(), shape=square_matrix_shapes, bound=10))
 def test_eigvals(x):
     res = linalg.eigvals(x)
     expected_dtype = dh.complex_dtype_for(x.dtype)