The following improvements were made to vtkDepthSortPolyData:

* transfrom GetCell to GetCellPoints. Building the cells one by one
  is expensive and we only need points to determine the depth. We
  can also then access the points in place.
* transform qsort to std::sort. Comparisons in std::sort get inlined
  so std::sort is much faster. I also reduced the overhead of swaps
  by using a functor so that only cell id is swapped.
* use templates to deal with point types instead of going through
  virtual GetValue API that converts them to double.
* restructure depth computations so that they can be vectorized by
  the compiler.
* I added a Cxx test to improve test coverage of the various sorting
  modes supported by this class.

The following improvements were made to vtkPolyData and vtkCellTypes:

* move the non-virtual impementations of GetCellPoints to
  the header to improve compiler optimization.
* add non-virtual GetCell method and also make it inline
* fix an issue in cell types where the MaxId is not set
  correctly when you pass in pre-built arrays

In tests the improved depth sort is ~ 3 x faster when sorting by
first point and cell bounds, and ~ 2 x faster when sorting by
parametric center. The tests made use of a ~ 8.1M cell iso-surface
computed from a cosmology simulation, and gcc 4.9.2 compiled with
-Wall -Wextra -DNDEBUG -Ofast -march=native -mavx -ffast-math

The patch includes the changes requested in the code review:
transform malloc/free to new[]/delete[], and transform memset
to std::fill_n.