Removed dropout layers from models using it

After discussion with @htoyryla (jcjohnson#408 (comment)), I think that processing fragments with exactly same sequence without random variations could make overlapping parts more conforming to each other.
Or maybe not. Anyway, it doesn't seem to makes results worse, and then it's simply faster.

neural_style.lua

@@ -208,8 +208,12 @@ print("Processing image part #" .. fragment_counter .. " ([" .. fx .. ", " .. fy
       local layer = cnn:get(i)
       local name = layer.name
       local layer_type = torch.type(layer)
+      local is_dropout = (layer_type == 'nn.Dropout')
       local is_pooling = (layer_type == 'cudnn.SpatialMaxPooling' or layer_type == 'nn.SpatialMaxPooling')
-      if is_pooling and params.pooling == 'avg' then
+      if is_dropout then
+        local msg = 'Removing dropout at layer %d'
+        print(string.format(msg, i))
+      elseif is_pooling and params.pooling == 'avg' then
         assert(layer.padW == 0 and layer.padH == 0)
         local kW, kH = layer.kW, layer.kH
         local dW, dH = layer.dW, layer.dH