Add Flatten layer to architectures + make flatten optional in embedders (#483)

Flatten layer required for embedders that mix conv and dense
(Cherry picking from #478)

rl_coach/architectures/layers.py

@@ -76,3 +76,11 @@ class NoisyNetDense(object):
 
     def __str__(self):
         return "Noisy Dense (num outputs = {})".format(self.units)
+
+
+class Flatten(object):
+    """
+    Base class for framework specific flatten layer (used to convert 3D convolution output to 1D dense input)
+    """
+    def __str__(self):
+        return "Flatten"