Fix Learner2D state rollback in BalancingLearner ask(tell_pending=False) (#492)

* Initial plan

* Fix Learner2D restore snapshot mutation in balancing asks

* Use shallow copies instead of deepcopy in Learner2D.__getstate__

Key-level mutations (stack pops/inserts) are all that restore() needs to
roll back; values are never mutated in place. A deepcopy of the full
data dict would add O(npoints) allocation to every
ask(tell_pending=False) call and to pickling.

---------

Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: Bas Nijholt <bas@nijho.lt>

Author: Copilot

adaptive/learner/learner2D.py

@@ -897,8 +897,8 @@ def __getstate__(self):
             cloudpickle.dumps(self.function),
             self.bounds,
             self.loss_per_triangle,
-            self._stack,
-            self._get_data(),
+            self._stack.copy(),
+            self._get_data().copy(),
         )
 
     def __setstate__(self, state):

adaptive/tests/test_balancing_learner.py

@@ -2,7 +2,7 @@
 
 import pytest
 
-from adaptive.learner import BalancingLearner, Learner1D
+from adaptive.learner import BalancingLearner, Learner1D, Learner2D
 from adaptive.runner import simple
 
 strategies = ["loss", "loss_improvements", "npoints", "cycle"]
@@ -51,6 +51,18 @@ def test_ask_0(strategy):
     assert len(points) == 0
 
 
+def test_ask_without_pending_restores_learner2d_state():
+    learner = Learner2D(lambda xy: xy[0] + xy[1], bounds=((-1, 1), (-1, 1)))
+    initial_stack = list(learner._stack.items())
+    initial_data = learner.data.copy()
+
+    balancing_learner = BalancingLearner([learner])
+    balancing_learner.ask(1, tell_pending=False)
+
+    assert list(learner._stack.items()) == initial_stack
+    assert learner.data == initial_data
+
+
 @pytest.mark.parametrize(
     "strategy, goal_type, goal",
     [