https://github.com/moosotc/llpp/issues/75

Looks like it's possible to make "progressive zoom" "nicer" by taking
already rendered tiles from the cache and scale them via GL in place
of not-yet-fully-rendered(by mupdf) ones, thus avoiding place holders
and providing more visually pleasant experience.

There was some proof of concept code to do that, but the
implementation got lost, then build system took precedence, and
finally - "writers block".

As a first step following algorithm should work:

 If tile isn't ready find all suitable tiles (same color-space etc etc)
 and if "scaled" version has non-empty intersection with a tile that
 is missing draw (scaled by GL version of) that instead.

 This, in my mind, potentially implies quite a bit of overdraw,
 simplest solution appears to be to just use the dimensional
 "closeness" as a Z value and enable and use depth testing when
 rendering the cached and scaled tiles during this pass (instead of
 some form of "painters algorithm" or something more elaborate)
