Fixes for having NumpySharedRunningStats syncing on multi-node (#139)

1. Having the standard checkpoint prefix in order for the data store to grab it, and sync it to S3.
2. Removing the reference to Redis so that it won't try to pickle that in.
3. Enable restoring a checkpoint into a single-worker run, which was saved by a single-node-multiple-worker run.

rl_coach/memories/non_episodic/differentiable_neural_dictionary.py

@@ -269,14 +269,19 @@ class QDND(object):
 
 
 def load_dnd(model_dir):
-    max_id = 0
+    latest_checkpoint_id = -1
+    latest_checkpoint = ''
+    # get all checkpoint files
+    for fname in os.listdir(model_dir):
+        path = os.path.join(model_dir, fname)
+        if os.path.isdir(path) or fname.split('.')[-1] != 'srs':
+            continue
+        checkpoint_id = int(fname.split('_')[0])
+        if checkpoint_id > latest_checkpoint_id:
+            latest_checkpoint = fname
+            latest_checkpoint_id = checkpoint_id
 
-    for f in [s for s in os.listdir(model_dir) if s.endswith('.dnd')]:
-        if int(f.split('.')[0]) > max_id:
-            max_id = int(f.split('.')[0])
-
-    model_path = str(max_id) + '.dnd'
-    with open(os.path.join(model_dir, model_path), 'rb') as f:
+    with open(os.path.join(model_dir, str(latest_checkpoint)), 'rb') as f:
         DND = pickle.load(f)
 
         for a in range(DND.num_actions):