upgrade grafana image to 4.1.1

hack/cluster-monitoring: hide expected errors and print meaning

hack/*: use kubectl apply where possible

There is an issue with kubectl apply and third party resources, so for
the time being we use create for TPRs, but apply for all manifests that
it can handle.

Don't pass $KUBECONFIG to kubectl --kubeconfig flag

The $KUBECONFIG variable is like $PATH, and supports multiple files
separated by `:`, but --kubeconfig only takes a single file as a value.
Since kubectl picks up the $KUBECONFIG variable already, don't pass it
to kubectl.

Also, use --namespace instead of -n to support older kubectl versions.

*: cleanup and unify minikube and self-hosted discovery

Update configs/alerts

Recently the ConfigMap for the prometheus config was changed, but not in
the assets, where it should have happened.

bootkube starts two replicas of the scheduler and controllers manager

correctly discover the Alertmanager cluster and ServiceMonitors

use prometheus-operator v0.1.0

v0.1.0 introduced handling Alertmanager clusters.

Fixed docker image tag for "quay.io/prometheus/node-exporter". 

The tag did not exist for "prometheus/node-exporter". So I have changed it to a valid value of "v0.13.0"

Update manifests to latest version of Prometheus Operator

The ports reflect the upstream kube-dns manifests of bootkube and
kubernetes/kubernetes.

The latest version of the Prometheus Operator requires Prometheus
>=v1.4.0 for the Alertmanger discovery feature.

These annotations made sense in pre v1.3.0 Prometheus releases, however,
with >=v1.3.0 and the Prometheus Operator these annotations are more
confusing then helpful.

Spleling

Fix broken kubernetes.rules

Fix always firing alerts

Scrape Prometheus service itself

This replaces the cluster overview dashboards with new dashboards
for a cluster-wide overview and inspecting single nodes.

Extract prometheus.yaml into assets