perf: eliminate allocation overhead in Configuration.add()/discard()

Mutate the OrderedSet in place and write back via the setter instead of
creating a new OrderedSet on every call. This removes the ~4-5% overhead
introduced by the persistence bugfix while preserving correctness for
both plain-attribute and deserializing-property models.

Author: fgmacedo

statemachine/configuration.py

@@ -106,23 +106,22 @@ def add(self, state: "State"):
         if csv is None:
             self.value = state.value
         elif isinstance(csv, MutableSet):
-            new = OrderedSet(csv)
-            new.add(state.value)
-            self.value = new
+            csv.add(state.value)
+            self.value = csv
         else:
             self.value = OrderedSet([csv, state.value])
 
     def discard(self, state: "State"):
         """Remove *state* from the configuration, normalizing back to scalar."""
         csv = self.value
         if isinstance(csv, MutableSet):
-            new = OrderedSet(v for v in csv if v != state.value)
-            if len(new) == 0:
+            csv.discard(state.value)
+            if len(csv) == 0:
                 self.value = None
-            elif len(new) == 1:
-                self.value = next(iter(new))
+            elif len(csv) == 1:
+                self.value = next(iter(csv))
             else:
-                self.value = new
+                self.value = csv
         elif csv == state.value:
             self.value = None