Fix TextureAtlas issues

[einarf]

The current TextureAtlas is weakly referecing textures supporting python garbage collection. The plan was to completely remove this due to the complexity it adds. I think we've maanged to figure out the remaining issues.

Fixes:

• Textures are now stored using the Texture.atlas_name meaning we only store textures with different hash-orientation combo. It doesn't care what hitbox algo the texture has. This mean the number of textures in the atlas are always the same as the uv data simplifying things.
• Resizing the atlas should no longer clear the image counter
• Rebuilding the atlas should now clear the image counter
• The atlas will now smartly rebuild itself if images have been removed before attempting to resize
• Added an initial ABC class for atlases
• Added sanity check in image ref counter raising error on underflow
• Fixed the experimenal atlas load/save code
• Added various comments to make it easier to inspect the code later