This required completely reworking startup sequence in `autocluster.erl`
without making the existing dozen step sequence into an incomprehensible
mess. Now steps are more independent and are executed in order with
function that provides something like State + Either monads.

Boot process is now split into 3 stages:
- First stage roughly corresponds to original implementation
- Second step is responsible for registering node in the backend. Also
  TTL timers are now started there, instead of spreading
  `-rabbit_boot_steps` through backend modules. This makes
  `autocluster.erl` the single source of truth about what is actually
  happening during startup.
- Third stage is responsible for releasing startup lock. We need it
  because of `ignore` failure mode - we want the lock to be released
  even if some prior steps failed.

During this rework it became evident that explicit mnesia:reset/0 is not
needed (more thorough explanation is in comments for
`autocluster:maybe_cluster/1`).

Startup locking support should work the same way for every backend, but
for now the only working implementation is for `etcd`. It even has some
property-based tests =)
Other backend fall back to original random delay behaviour.

Also it looks like `dialyzer` ignores type specs in comments - because
some of types there were definitely wrong, and converting to them to
`-spec` caused a lot of barfing from `dialyzer`.