Fix normals

[LukasKalbertodt]

The normal calculation is not usable with a very detailed mesh (like mandelbulb at 5 iterations or more). The whole calculation depends critically on the sampling offset. Right now it's at 0.01 for no apparent reason. Lowering this value makes the normal match the "real" normal more closely. but this might not be expected as the users sees a rather coarse mesh approximation of the shape anyway. A value of 1.0 would probably match the normal of the shown mesh exactly.

My point: we need to test different values or methods to determine the normal. Related to #6.

[LukasKalbertodt]

The value was adjusted to 0.7 in #34. This looks pretty ok.

[LukasKalbertodt]

Right now, I think the best way is to calculate the normals based on the resulting mesh.