Original rework of caches during undo/redo (see D8183) had a very bad
flaw hidden in it: using the key of a ghash as source of data.

While this was effectively working then (cache pointer itself being part
of the key, and said cache pointers not being cleared on file write),
this is a general very bad way to do things.

Now that cache pointers are more and more cleared on file write (as part
of clearing runtime-data to reduce false-positives when checking if an
ID has changed or not), this has to be fixed properly by:

• Not storing the cache pointer itself in the IDCacheKey.
• In undo context, in readfile code trying to preserve caches, store the cache pointers as values of the mapping, together with the usages counter

The first change potentially affects all usages of
BKE_idtype_id_foreach_cache, but in practice this code is only used by
memfile reading code (i.e. undo) currently.

Related to T97015.