Rewrite iter_subtrees to more efficient version

Using OrderedDict provides these properties:
- given subtree is yielded only once since it's stored in OrderedDict
only once (even though it may be put there multiple times) so no need
to double check if subtree was already seen
- order of iteration is preserved as it was previously when subtrees
to iterate over were stored in a list

lark/tree.py

@@ -4,6 +4,7 @@ except ImportError:
     pass
 
 from copy import deepcopy
+from collections import OrderedDict
 
 
 ###{standalone
@@ -58,25 +59,17 @@ class Tree(object):
         return hash((self.data, tuple(self.children)))
 
     def iter_subtrees(self):
-        # TODO: Re-write as a more efficient version
-
-        visited = set()
-        q = [self]
-
-        l = []
-        while q:
-            subtree = q.pop()
-            l.append( subtree )
-            if id(subtree) in visited:
-                continue    # already been here from another branch
-            visited.add(id(subtree))
-            q += [c for c in subtree.children if isinstance(c, Tree)]
-
-        seen = set()
-        for x in reversed(l):
-            if id(x) not in seen:
-                yield x
-                seen.add(id(x))
+        queue = [self]
+        subtrees = OrderedDict()
+        for subtree in queue:
+            if id(subtree) in subtrees:
+                continue
+            subtrees[id(subtree)] = subtree
+            queue += [c for c in reversed(subtree.children) if isinstance(c, Tree)]
+
+        del queue
+        for subtree in reversed(list(subtrees.values())):
+            yield subtree
 
     def find_pred(self, pred):
         "Find all nodes where pred(tree) == True"