We must call on_new_interval() on any interval change *and* on the
creation of the PG. Currently we call it from PG::init() and
PG::start_peering_interval(). However, PG::split_into() did not
do so for the child PG, which meant that the new child feature
bits were not properly initialized and the bitwise/nibblewise
debug bit was not correctly set. That, in turn, could lead to
various misbehaviors, the most obvious of which is scrub errors
due to the sort order mismatch.
Fixes: #13962 Signed-off-by: Sage Weil <sage@redhat.com>