This fixes a couple of bug:
 - Bug #3740 happens because `Repeater::model_set_row_data` did not use
   the inner.offset to get the instance (So that's the `val == 106` part
   of the test)
 - But I went ahead and also tested what happenned if you changed the
   model from the "outside" using the `model[i] = `, and that was not
   implemented, hence the move of the code from
   `Repeater::model_set_row_data` to `RepeaterTracker::row_changed`,
   That does need Pin though, so Pin was added everywhere
 - C++ is not affected by bug #3740, because because the C++ listview
   don't do the "allocate only visible" optimization. But the
   "val == 1106" part of the test would fail so this patch also moces
   the update from  `model_set_row_data` to `row_changed`.
   But following that we don't set the state as Dirty in `row_changed`,
   the write_to_model test started filling  because the `row_added`
   function was missing an update of the index on every further items

The change in the interpreter prevent a borrow_mut from causing trouble.
The Value really don't need to be in a RefCell anyway

Fix #3740