Simplified cleanup_cursor() implementation

I hit this performance bug when I cloneda 40G NTFS partition. The actual
copy was done in under 11 minutes. After that I was shocked to find that
gparted would spend over 12 minutes in cleanup_cursor, pegging a CPU
core. (On a quad core desktop...)

Simply replacing the ustring with std::string would reduce the time to
about 1.5 minutes. Still bad. Also, I didn't want to lose UTF8
awareness.

So I rewrote the algorithm in 'streaming mode'. This has the (potential)
drawback that locale conversions are done, but performs well and
- IMHO - is a lot more readable.

On a minor note: this implementation correctly handles backspaces at the
start of a line.