etcd v3 support

[mumoshu]

ref kubernetes/kubernetes#22448

Notes:

• Etcd v3 is able to serve requests for both the v2 and v3 data
• Running etcd v3 against the v2 data dir doesn't automatically upgrade the data dir to the v3 format
• Keeping to use v2 api against etcd v3 only updates the v2 data stored in etcd
• Though it might be out of scope of kube-aws, upgrading of etcd to v3 is documented in the etcd doc
• The whole migration process is described in the CoreOS blog
• etcdctl should have an env var ETCDCTL_API=2 or ETCDCTL_API=3 provided to talk with an etcd cluster in API v2 or v3 respectively.
  • See https://github.com/coreos/etcd/blob/master/Documentation/upgrades/upgrade_3_0.md#further-considerations
• kube-apiserver talks with an etcd cluster in API v3 when --storage-backend etcd3 is provided.
• Etcd3 unit could be configured like https://github.com/coreos/matchbox/blob/master/examples/ignition/etcd3.yaml

[mumoshu]

etcd3 seems to simplify the process of the disaster recovery of etcd clusters, and hence its implementation in kube-aws.
See #27

[mumoshu]

Verified etcd3 does allow each member to recover from its snapshot "when in an initial cluster bootstrapping process", where etcd2 does not.
We still need cluster runtime reconfiguration, which I wanted to avoid in the first place, to recover from a partial permanent failures of nodes up to N/2 but this is the same as etcd2.

See here and here in my prototype which runs inside coreos vagrant machine

[mumoshu]

Moving it to v0.9.6 because it seems to require kubernetes 1.6 to pass all the k8s E2E tests.