Improve performance with large cubes

[astrofrog]

There are several optimizations we can make when dealing with large cubes:

• Change the relative step size (make it an option to do this automatically when panning/zooming?)
• Make it so the limits for the image are applied after the maximum intensity has been found in the shader, not in the Python code beforehand
• Don't copy the data to get rid of the NaNs but see if we can deal with it better in the OpenGL code

I'm working on this.

[astrofrog]

The volume rendering now downsamples on the fly while interacting with it, and I've also sped up the code in various places.