Rounding Safe Reverse McCormick Relaxations

[mewilhel]

Our basic MC objects are not numerically safe. As a consequence, the reverse McCormick relaxations can produce erroneous results. This was a noted issue in the original Wechsung paper. A solution to this consists of defining a corrected rounded relaxation types NSSafe, MVSafe, DiffSafe in McCormick then using these to define backwards passes (that said, this is a lot of work, so we'll need to kick this down the rado to McCormick 2.0 or 3.0).