Fix issue exporting graph with multi-GPU

[harperj]

Our multi-GPU training had a regression such that freezing the
graph was broken. This change fixes that issue by making a few
changes:

• Removes the top level "tower" variable scope added by multi-GPU
  so that the output nodes have correct names
• Removes the use of "freeze_graph" and updates the tensorflow to
  barracuda converter to accept GraphDef objects
• Adds the "auto reuse" to network layers which require them

[CLAassistant]

Thank you for your submission, we really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ harperj
❌ harper-u3d
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[harperj]

Updated this PR to remove changes to tensorflow_to_barracuda.py by simply writing out the frozen graph definition using the same approach as the freeze_graph function from Tensorflow. I've tested that Barracuda inference on a saved .nn model and also restarting training using --load still work as expected.