Implements part of #535

Some things we can talk:

• We can omit state="stale" like labels and just give the state value to Gauge (IMO current impl is simpler) although I saw below in CONTRIBUTING.md I didn't see many magic numbers in the metrics and labels thus added human readable names for them.

For example a proc file contains the magic number 42 as some identifier, the Node Exporter should expose it as it is and not keep a mapping in code to make this human readable.

• For code quality I can split Update function to make it more readable (again IMO it's simple enough to do it in one function)

This pr adds below metrics

# HELP node_arp_states ARP states by device
# TYPE node_arp_states gauge
node_arp_states{device="eth0",state="stale"} 1
node_arp_states{device="eth0",state="failed"} 2
# ...

Old comments

I changed the design in second commit due to not exploding cardinality but keep the comments if maintainers wants that design.

Iff you want to have metrics like below, can continue to read:

node_arp_states{device="eth0",state="stale", ip="1.1.1.1"} 1
node_arp_states{device="eth0",state="stale", ip="1.1.1.2"} 1
node_arp_states{device="eth0",state="failed", ip="1.1.1.3"} 1

Some things we can talk:

• Can remove neighborState struct and just use string slice, but this is easier to follow IMO
• This change may create many metrics, to make a estimation Default values of arp cache size are: min 128 and max 1024 therefore we may except up to 1024*count(net_int) as max. But in some "fine tuning"s I saw people generally increase this although it's limited by the size of you subnet, one can have many machines in subnet. Maybe I should make a counter for each state?

@SuperQ it's ready for your review, when you have time 🙏🏻