Fix a problem with the mapper that breaks volume rendering when
the active vtkCamera has a non-identity ModelTransformMatrix.

Since the eye position is needed in data space, and the camera's
ModelViewTransform can affect that, combine camera and volume
matrices to infer the eye position for each volume.  Also, this
can be computed on the cpu, avoiding a matrix/matrix multiply and
and handful of matrix/vector multiplies in the fragment shader.