> where each part could be GCed separately

That would mean disallowing cyclical references across different GC heaps; non-cyclical references would just create additional GC roots. It would be a step away from totally automated memory management and towards something closer to "smart" reference counting.