undo grouping should be more intelligent

[ossi]

as a followup to #27, i'm creating a new ticket as requested.

undo grouping should roughly predict what the user probably wants to undo at once, without grouping too much. i suggest an action/time/space based grouping:

• if the user switches to another "action sequence" (inserting/overwriting, deleting, navigating, maybe more), he certainly wants it separated from the previous sequence
• if he makes a longer break while doing things, he probably expects it when undoing as well. what "longer" means is very subjective; a simple adaptive algorithm might make sense
• small moves are merged, while big ones aren't. i'm not even sure what the criteria should be here. maybe moves should be generally merged and we should only depend on the other two "break conditions".