Improve `numpy.approx` array-scalar comparisons

So that `self.expected` in ApproxNumpy is always a numpy array.

_pytest/python_api.py

@@ -87,14 +87,15 @@ class ApproxNumpy(ApproxBase):
     def __eq__(self, actual):
         import numpy as np
 
+        # self.expected is supposed to always be an array here
+
         if not np.isscalar(actual):
             try:
                 actual = np.asarray(actual)
             except:  # noqa
                 raise TypeError("cannot compare '{0}' to numpy.ndarray".format(actual))
 
-        if (not np.isscalar(self.expected) and not np.isscalar(actual)
-                and actual.shape != self.expected.shape):
+        if not np.isscalar(actual) and actual.shape != self.expected.shape:
             return False
 
         return ApproxBase.__eq__(self, actual)
@@ -102,16 +103,11 @@ class ApproxNumpy(ApproxBase):
     def _yield_comparisons(self, actual):
         import numpy as np
 
-        # For both `actual` and `self.expected`, they can independently be
-        # either a `numpy.array` or a scalar (but both can't be scalar,
-        # in this case an `ApproxScalar` is used).
-        # They are treated in `__eq__` before being passed to
-        # `ApproxBase.__eq__`, which is the only method that calls this one.
+        # `actual` can either be a numpy array or a scalar, it is treated in
+        # `__eq__` before being passed to `ApproxBase.__eq__`, which is the
+        # only method that calls this one.
 
-        if np.isscalar(self.expected):
-            for i in np.ndindex(actual.shape):
-                yield np.asscalar(actual[i]), self.expected
-        elif np.isscalar(actual):
+        if np.isscalar(actual):
             for i in np.ndindex(self.expected.shape):
                 yield actual, np.asscalar(self.expected[i])
         else:
@@ -202,7 +198,7 @@ class ApproxScalar(ApproxBase):
         the pre-specified tolerance.
         """
         if _is_numpy_array(actual):
-            return actual == ApproxNumpy(self.expected, self.abs, self.rel, self.nan_ok)
+            return ApproxNumpy(actual, self.abs, self.rel, self.nan_ok) == self.expected
 
         # Short-circuit exact equality.
         if actual == self.expected: