Since ZoL frees ARC under memory pressure, let's not
count it as "used" but instead as "free" memory.

node_memory_MemAvailable_bytes is a better estimator than the sum I used before
says some Prometheus documentation.  It is also almost the same, but reads nicer.

This should fix the current alerts for our RAID arrays.
It's only "should" because I can't test it proper without said
RAID arrays in the dev or staging machines.

- Also alert on "negative" (== receiving) errors
- Use 'rate' so we can get out of the reporting if situation normalizes

Grafana 8 recommends these for added performance and capabilities.

Fixes privatestorageops#358.

This changes the Throughput graph to be relative, changes its unit
to Bit/s, and fixes the labels of the Network pkt/s graph.

Fixes privatestorageops#24.

Only dashboards in this dir

Fixes privatestorageops#349.

Make everything more consistent and generally more nice.
I am a bit proud of this now.
Let's hope it holds up.

Rationale: These are servers, they should be transmitting mostly.
But really it's rather arbitrary.

- Add node exporter to nodes
- Make the old nixpkgs 19.09 node exporter work (amend config)
- Fix a typo (??) and exclude monitoringvpn in network traffic graphs to prevent double counting