With certain malformed jpegs the call to jpeg_finish_decompress
caused a long jump back to stop_loading. This jump overwrote the
value of the surface pointer which should be NULL which in turn 
caused a cairo assert when trying to get a reference to a surface
with ref_count = 0.

Also moved the cleanup of in_buffer out of the !finished scope, 
as it seems this would otherwise be a potential leak.