Optimizations based on Druid callgrind dump sent by Manqi.

* TGLBoundingBox:
  Use TGLVertex3[8] instead of std::vector<> for member that always
  has 8 elements. Caused pointless allocations and fragmentation.

* TGLScene:
  In CalcBoundingBox() avoid creating a bounding box temporary in a
  loop over all elements -- use const& instead.

* TGLCamera:
  Fix for changes in TGLBoundingBox.


git-svn-id: http://root.cern.ch/svn/root/trunk@37397 27541ba8-7e3a-0410-8455-c3a389f83636