Fix use of removed array rules in test (#952)

This fixes the failing test and two performance warnings raised in the
tests. Like @henryiii suggested, these changes were needed in one
particular test which relied on broadcasting rules (in this case they
were narrowing rules) that have been removed from numpy.

Previously, it was possible to do this
```
a = np.zeroes(2)
b = np.ones(1)
a[0] += b # add sequence of length 1 to scalar
```

src/iminuit/util.py

@@ -236,7 +236,7 @@ def _set(self, i: int, arg: UserBound) -> None:
         state.set_error(i, err)
 
 
-def _normalize_limit(lim: UserBound) -> Tuple[float, float]:
+def _normalize_limit(lim: Optional[Iterable]) -> Tuple[float, float]:
     if lim is None:
         return (-np.inf, np.inf)
     a, b = lim

tests/test_cost.py

@@ -1164,7 +1164,7 @@ def grad2(x, b, a):
         return g
 
     def model3(x, c):
-        return c
+        return c * np.ones_like(x)
 
     def grad3(x, c):
         return np.ones((1, len(x)))
@@ -1201,7 +1201,7 @@ def grad3(x, c):
         a = 2
         b = 3
         ref = np.zeros(2)
-        ref[0] += lsq1.grad(a)
+        ref[0] += lsq1.grad(a)[0]
         ref[[1, 0]] += lsq2.grad(b, a)
         assert_allclose(lsq12.grad(a, b), ref)
 
@@ -1214,9 +1214,9 @@ def grad3(x, c):
         a = 2
         b = 3
         ref = np.zeros(2)
-        ref[0] += lsq1.grad(a)
+        ref[0] += lsq1.grad(a)[0]
         ref[[1, 0]] += lsq2.grad(b, a)
-        ref[0] += lsq1.grad(a)
+        ref[0] += lsq1.grad(a)[0]
         assert_allclose(lsq121.grad(a, b), ref)
 
     lsq312 = lsq3 + lsq12
@@ -1229,8 +1229,8 @@ def grad3(x, c):
         b = 3
         c = 4
         ref = np.zeros(3)
-        ref[0] += lsq3.grad(c)
-        ref[1] += lsq1.grad(a)
+        ref[0] += lsq3.grad(c)[0]
+        ref[1] += lsq1.grad(a)[0]
         ref[[2, 1]] += lsq2.grad(b, a)
         assert_allclose(lsq312.grad(c, a, b), ref)
 

tests/test_minimize.py

@@ -63,9 +63,9 @@ def rosen(par):
 
 
 def test_disp(capsys):
-    minimize(lambda x: x**2, 0)
+    minimize(lambda x: np.sum(x**2), 0)
     assert capsys.readouterr()[0] == ""
-    minimize(lambda x: x**2, 0, options={"disp": True})
+    minimize(lambda x: np.sum(x**2), 0, options={"disp": True})
     assert capsys.readouterr()[0] != ""
 
 
@@ -115,7 +115,7 @@ class Fcn:
 
         def __call__(self, x):
             self.n += 1
-            return x**2 + 1e-2 * (self.n % 3)
+            return np.sum(x**2 + 1e-2 * (self.n % 3))
 
     r = minimize(Fcn(), [1], options={"maxfun": 100000000})
     assert not r.success