fix: dont mistake non-key access with foreign key

[patricebender]

Consider this two path expressions:
classrooms.classroom.ID, classrooms.classroom.name

While the first path could be optimized to classrooms.classroom_ID, the second path is not a key access and hence must be transformed to classroom.name, where classroom is the nested join node, refer to this example:

    const query = CQL`SELECT from Pupils {
      ID
    } group by classrooms.classroom.ID, classrooms.classroom.name`
    const expected = CQL`SELECT from Pupils as Pupils
                        left join ClassroomsPupils as classrooms
                          on classrooms.pupil_ID = Pupils.ID
                        left join Classrooms as classroom
                          on classroom.ID = classrooms.classroom_ID
                        {
                          Pupils.ID
                        } group by classroom.ID, classroom.name`

As both paths navigate through the same join nodes, we do not (yet?) optimize the first path to the foreign key, but access the ID in the target of classroom. This is also the behavior of the cds-compiler and should not matter too much as the join node is produced through the second path anyway:

Pupils
  └── classrooms
        └── classroom
              ├── ID
              └── name

The problem was, that when we found the second path, we have already seen this path up to the n - 1 step before, and only the reference for the first path was adjusted to not be a foreign key access anymore.

fixes cap/issue#16001