Blender Git Loki

undoexp: initial implementation of ID swapping.

Goal is to reuse same address also for changed IDs during an undo step.
This will allow for a given ID to almost always keep its same address
along a whole editing session, reducing a lot useless updates required
by current memfile undo.

Note that this commit has a lot of things to be investigated & fixed
still, at the very least:
* Refcounting is not really taken care of yet. This is fine for now
since we remap everything still, but at some point plan is to not remap
(liblink) reused unchanged IDs at all...
* We keep the double libmap for now, getting rid of it requires further
investigations (especially in some corner cases).
* 'UI' IDs (WM, screen and WS) are likely troublemakers given their
current weird specific handling. This might actually be seriously
simplified with this new approach?
* Since even changed IDs keep the same address, we are going to have to
improve a lot the `id->recalc` handling - unless we accept a
brute-force complete depsgraph update of the changed IDs.

Last point is especially interesting, as with current code, updates do
happen as expected most of the time (proper updates are missing
sometimes). This means that even undo moving of a very highly
modifier-subdivided object can be done "instantaneously" as that object
updates does not requires a geometry update.

Handling it properly will likely require to store the 'future' update
flag (as we do currently with the future unchanged status). Not sure
we'll catch all cases though, if that goes too far we may just force
full update as a first step. :/