Snapshot usage makes qcow2 image unusable due to large tables

To reproduce with 0.9.1 and svn:
- Create a 20G (or some size much greater than system RAM) qcow2 image
- Inside VM, install some OS, formatting whole drive
- Create snapshot with savevm
- Inside VM, reformat and reinstall OS
- Create snapshot with savevm
[...]

Eventually, qemu crashes, then neither qemu-img nor qemu can open the image because memory is exhausted.  The reason is that the whole refcount_table is loaded into memory, and this refcount_table has now become much bigger than the size of memory.

The refcount_table really needs to be loaded and used in fixed size chunks to avoid this problem.

Alternatively, there needs to be a way to "rollback" a snapshot without loading the whole disk image normally, so that a snapshot which has made the image unusable in this way can be reversed.