Hi Tobias,
1. As José Aliste explained it (https://sourceforge.net/tracker/?func=detail&aid=2796407&group_id=204144&atid=988431) evince caches every single page you read in the document, and renders it asynchronously. It is way faster, but uses much more memory (it often happens to me when having many documents open at the same time that evince takes up to 700-800Mb RAM, and 150 Mb for only one file). In particular, the more you zoom, the more it takes memory. I guess this is why evince only lets you zoom to 400%.
So this is a matter of choice (and of programming time!). When EvView will be out, and José will have implemented the preview with it, it should be much faster.

2. I think this is the normal behaviour. After you filed this, I tried different things to prevent this, but without success... It is of course related to (point 1.) the fact that the rendering is not done asynchronously, and that PyGTK does only one GUI thing at a time.

3. This is solved in svn revision 371, and should thus be in 0.2 final.