- added allow_gc option to PerformLinker and OpWiseCLinker, which results in
  temporary results being freed as soon as possible.  This can have a negative
  impact on performance (we have to re-allocate outputs) but allows us to
  evaluate graphs piecewise, even if not all temporaries fit in memory at
  once.

- modified Function so that internal references to `required` input arguments
  and computed output arguments are removed after a call.  This means that
  these values are not stored by pickle.

- modified MergeOptimizer to use a more efficient algorithm.  The previous one
  was at best quadratic in depth, and at worst exponential in depth.  The new
  one should be worst-case linear.