Set up GPU device if available

[ldx]

Two main changes:

• Build AMI with a relatively recent NVIDIA driver and toolchain.
• Set up units for GPU access if an NVIDIA GPU is detected on the host.

I also built a dev AMI, but I have not promoted it to prod yet. Probably we want to pin our excessively large customer base to the current stable itzo release and prod AMI first.

To test it, have server.yml use the milpadev AMI and set the default instance type to e.g. p2.xlarge. Then create a pod using e.g. tensorflow-gpu:

kubectl run tf --image=tensorflow/tensorflow:latest-gpu --overrides='{"apiVersion": "apps/v1", "spec": {"template": {"metadata": {"annotations": {"kubernetes.io/target-runtime":"kiyot"}}, "spec": {"nodeSelector": {"kubernetes.io/role": "milpa-worker"}}}}}' --command=true -- sleep infinity

Once it's running, exec into the pod, and check that tensorflow is able to detect and use the GPU:

python -c "import tensorflow as tf; sess = tf.Session(config=tf.ConfigProto(log_device_placement=True))"

Next step is to have kiyot request GPU(s) when the pod sets a GPU resource limit, and implement our deviceplugin for the kubelet that exposes GPU capabilities.

[justnoise]

Great job getting this working!