Reduce overhead in Executor critical section.

This change changes two incidences of `InlinedVector<TaggedNode>::push_back()` (of a newly constructed TaggedNode) into `InlinedVector<TaggedNode>::emplace_back()`. Since one of these instances happens in `ExecutorState::FrameState::ActivateNodes()`, which is called when the FrameState mutex is held, the serial performance improvement should compound with reduced contention on that mutex.

PiperOrigin-RevId: 211094227