FIX: Fix unit example so that we can unpin numpy>2.1

Closes matplotlib#28780.

The underlying problem is that operations on numpy scalars try to
eagerly convert the other operand to an array. As a result `scalar = np
.float64
(2); scalar * radians` would result in a numpy scalar. But we don't want
 that.
Instead we enforce `radians.__rmul__(scalar)` by giving the unit a
higher
`__array_priority__`. See also https://github
.com/numpy/numpy/issues/17650.

I haven't found any specific change notes on this in numpy 2.1.
Interestingly, the full story is even more complex. Also for numpy<2.1
`radians.__rmul__(scalar)` is not called, but there seems another
mechanism through __array__ and __array_warp__ catching back in so that
the result is again a TaggedValue. But I have not fully investigated why
 it worked previously. In fact, we want the solution here with going
 through __rmul__, and that works for all numpy versions.

`

Author: timhoffm

.circleci/config.yml

@@ -98,7 +98,7 @@ commands:
     parameters:
       numpy_version:
         type: string
-        default: "~=2.0.0"
+        default: ""
     steps:
       - run:
           name: Install Python dependencies

environment.yml

@@ -18,7 +18,7 @@ dependencies:
   - kiwisolver>=1.3.1
   - pybind11>=2.13.2
   - meson-python>=0.13.1
-  - numpy<2.1
+  - numpy
   - pillow>=9
   - pkg-config
   - pygobject

galleries/examples/units/basic_units.py

@@ -193,6 +193,11 @@ def get_unit(self):
 
 
 class BasicUnit:
+    # numpy scalars convert eager and np.float64(2) * BasicUnit('cm')
+    # would thus return a numpy scalar. To avoid this, we increase the
+    # priority of the BasicUnit.
+    __array_priority__ = np.float64(0).__array_priority__ + 1
+
     def __init__(self, name, fullname=None):
         self.name = name
         if fullname is None: