We need to be able to distinguish between reused IDs and newly read ones, and 'NEW' tag is not enough here, as it gets cleared in read code.
This will be used in next commit to inform depsgraph that some re-used old IDs need at least a COW refresh, as they are using ID pointers to newly read data.
undo experiments: actually write at leat one memchunck per ID.
Checked that before, but missed the fact that the flush was done after a whole ID type process in the main saving loop, now we actually do it after every ID when using mem undo file saving.
Besides fixing the stupid issue with lots of unchanged IDs being detected as changed when reading back the undo step, this should also help reduce undo memory usage in some cases (although I could not detect that with some quick test...).