Error spam from credential provider when not on GCE

[jbeda]

It looks like we always try and hit the metadata server even if we aren't on GCE. We then print an error out.

This should either be silent (assume after X attempts that we aren't on GCE?) or should only be activated with an option.

Email thread: https://groups.google.com/forum/#!topic/google-containers/w-ei5Xs6K0Y

[jbeda]

cc: @mattmoor

[mattmoor]

FYI, the way the credentialprovider logic is set up, providers are asked once per keyring creation whether they are enabled.
https://github.com/GoogleCloudPlatform/kubernetes/blob/master/pkg/credentialprovider/provider.go#L30

If they are enabled, then they are asked to provide their contribution to .dockercfg each time the lazy keyring is accessed.
https://github.com/GoogleCloudPlatform/kubernetes/blob/master/pkg/credentialprovider/keyring.go#L47

For efficiency, there is a caching provider, which composes with heavier providers and stores their .dockercfg contribution for a predetermined TTL.
https://github.com/GoogleCloudPlatform/kubernetes/blob/master/pkg/credentialprovider/provider.go#L50

What was happening here was that "Enabled()" for our GCE-metadata implementations was logging about errors fetching "http://metadata", where the entire "Enabled()" check is "err == nil".

[brendandburns]

I believe that this was fixed by #2674