PointOptimizerSummary option to avoid flattening (#2461)

This isn't actually enough as the point optimizer will flatten it (again). But this *does* make it so I can write my own PointOptimizer replacement that can respect it

xref #2406

cirq/circuits/optimization_pass.py

@@ -30,8 +30,11 @@
 class PointOptimizationSummary:
     """A description of a local optimization to perform."""
 
-    def __init__(self, clear_span: int, clear_qubits: Iterable['cirq.Qid'],
-                 new_operations: 'cirq.OP_TREE') -> None:
+    def __init__(self,
+                 clear_span: int,
+                 clear_qubits: Iterable['cirq.Qid'],
+                 new_operations: 'cirq.OP_TREE',
+                 preserve_moments: bool = False) -> None:
         """
         Args:
             clear_span: Defines the range of moments to affect. Specifically,
@@ -41,8 +44,17 @@ def __init__(self, clear_span: int, clear_qubits: Iterable['cirq.Qid'],
                 with each affected moment.
             new_operations: The operations to replace the cleared out
                 operations with.
+            preserve_moments: If set, `cirq.Moment` instances within
+                `new_operations` will be preserved exactly. Normally the
+                operations would be repacked to fit better into the
+                target space, which may move them between moments.
+                Please be advised that a PointOptimizer consuming this
+                summary will flatten operations no matter what,
+                see https://github.com/quantumlib/Cirq/issues/2406.
         """
-        self.new_operations = tuple(ops.flatten_op_tree(new_operations))
+        self.new_operations = tuple(
+            ops.flatten_op_tree(new_operations,
+                                preserve_moments=preserve_moments))
         self.clear_span = clear_span
         self.clear_qubits = tuple(clear_qubits)