- Failing handlers shouldn't cause the instrumented code to crash
- Subsequent handlers should still be executed
- There is no removal of broken handlers (like telemetry),
  because that would put the system in an inconsistent state if a
  handler fails intermittently.

The goal is to allow automatic initialisation/cleanup.

This way start and stop operations are more organised,
and the functionality is tested.

- The goal is to use them in the handler modules.
- label_key and label_value are a bit repetitive, but this way the specs
  are as strict as possible. If the list grows longer, we could make
  the type more generic instead of enumerating all values.

1. Support the 'all_hosts_are_global' option
2. Move prefix management from mongoose_metrics to eliminate
   the dependency on legacy code.
3. Reset already existing metrics on setup.
   The goal is to have fresh repeatable metric state if you restart
   mongoose_instrument in tests.
   There is no metric removal, because it doesn't seem to be needed.

More initialisation is needed because of host-type prefix handling.

1. Add type specs.
2. Use strings as metric names to avoid calling 'list_to_atom'.
   This doesn't seem to have a big performance penalty.
3. Reset already existing metrics on startup, just like for exometer.
4. Initialise counters with zero - previous initial value was 'undefined',
   resulting in a delay in initial rate metric calculation in Prometheus.

Metric name is now a string, not an atom.