Fix test_approx() when one of the arguments is a broadcasted multidimensional array.

`collect()` on broadcasted arrays doesn't preserve its shape, but instead creates a flat array. This breaks tests e.g. for most activation functions in NNlib:

```
test_rrule(Broadcast.broadcasted, NNlib.σ, rand(3, 4))
```

Author: dfdx

src/check_result.jl

@@ -142,8 +142,8 @@ function test_approx(actual::A, expected::E, msg=""; kwargs...) where {A,E}
     if _can_pass_early(actual, expected)
         @test true
     else
-        c_actual = collect(actual)
-        c_expected = collect(expected)
+        c_actual = collect(Broadcast.materialize(actual))
+        c_expected = collect(Broadcast.materialize(expected))
         if (c_actual isa A) && (c_expected isa E)  # prevent stack-overflow
             throw(MethodError, test_approx, (actual, expected))
         end

test/check_result.jl

@@ -36,6 +36,7 @@ end
 
             test_approx([1.0, 2.0], [1.0, 2.0])
             test_approx([[1.0], [2.0]], [[1.0], [2.0]])
+            test_approx(Broadcast.broadcasted(identity, [1.0 2.0; 3.0 4.0]), [1.0 2.0; 3.0 4.0])
 
             test_approx(@thunk(10 * 0.1 * [[1.0], [2.0]]), [[1.0], [2.0]])