Do not try to finish already completed phase

Do not try to finish already completed phase

Profiling shows that completion of substatement phases is the most
expensive operation we are performing. This stems from the fact
that inference is reactive and eager -- hence at we can attempt
to complete a particular phase against a statement multiple times.

Each time we will end up also walking all children (and their
children), attempting to complete the phase -- even when a particular
child has already successfully completed it.

This exposes a slight issue in our logic, as we have always assumed
child statements need to go through all the phases with us -- which
has not been true due to 'typedef' statements reuse, which would
end up being churned through phase complete (but always end with
EFFECTIVE_MODEL completed).

Fix both issues by checking whether the statement has executed
requested (or a subsequent) phase before going into actual phase
execution. This also requires that any substatements added to
a statement transition to the parent's completed phase before they
are made visible -- otherwise we could could end up with a statement
tree whose root has completed EFFECTIVE_MODEL, but not all of its
substatements have, triggering state violation assertions.

JIRA: YANGTOOLS-1082
Change-Id: Id6ed0d7feefc6d838c055e9690db69ae26633d95
Signed-off-by: Robert Varga <robert.varga@pantheon.tech>