This change identifies datasets that are far from their origins
(relative to their bounds), which can result in precision problems
when rendering. Problem datasets have their VBO modified and
transforms concatenated to the view/display transforms sent to the
shaders by the mapper to correct the problem.

OpenGL drivers often use reduced-precision point coordinates; if the
most significant bits are all used by a large translation then
renderings are often misleading and frame-to-frame differences
exaggerated. To counter this, this commit makes each VBO store shift
and scale vectors generated when CreateVBO is called. If the origin
is more than 10,000 bounding-box lengths away, transform the points
as they are written to the VBO (this condition forces a slow-path where
the VBO is generated, not copied directly from the vtkPoints array).
When this occurs, the VBO presents a flag that informs the mapper to
concatenate an inverse transform to the model-to-display and
model-to-viewport matrices. Since these transformations include the
camera projection and are done on the CPU with high precision, the
translation is factored out of the data when the camera is nearby.

This change also makes the VBO objects that the mappers own
publicly accessible so that applications can override the behavior,
either to manually set the shift and scale (as calling AppendVBO
multiple times cannot compute a proper shift and scale) or to
disable it should it cause issues.

Thanks to Ken Martin for suggesting this approach.