masks: don't let mask manager rows hand out stale module pointers

[da-phil]

masks: don't let mask manager rows hand out stale module pointers

Tree rows cache raw dt_iop_module_t pointers but outlive the modules,
which are freed on instance delete, image change and darkroom leave.
Selecting a group row then called flags() through a dangling pointer,
and deleting a shape crashed in dt_iop_is_hidden().

• _lib_masks_get_values(): only return a module still present in
  dev->iop; NULL attributes the edit to the mask manager
• _forms_structure_hash(): include the modules and their mask ids so a
  changed module set forces a rebuild
• invalidate the cached hash on image change and darkroom entry: the
  allocator may hand back the freed addresses and hash identical
• gui_update(): drop a pending selection module that died before the
  deferred update ran
• _tree_delete_shape(): stop clobbering the row's module with the one of
  the previous row

Speculatively fixes #22050

Disclaimer: this change was co-created with Claude.

[TurboGit]

Since mask manager will be removed in 5.8 and replaced by the new flexi masking maybe this is just not necessary.

[da-phil]

I guess it also depends if that gets into the way of the flexi masking development, like constantly producing merge conflicts. If not, I don't see why we should not make sure existing bugs are fixed until the module can be fully replaced by a new one. What do you think about that?