`ArrayPortalToIterators` broken on MSVC + CUDA

When compiling with MSVC, ArrayPortalToIterators uses a platform-specific checked-iterator class to wrap pointer accesses:

  // The MSVC compiler issues warnings when using raw pointer math when in
  // debug mode. To keep the compiler happy (and add some safety checks),
  // wrap the iterator in checked_array_iterator.
  using IteratorType = stdext::checked_array_iterator<IterType>;

This is incompatible with CUDA, since the checked_array_iterator lacks __host__ __device__ markup:

src\vtk-m\vtkm/Algorithms.h(143): warning #2979-D: calling a __host__ function from a __host__ __device__ function is not allowed
          detected during:
            instantiation of "IterT vtkm::UpperBound(IterT, IterT, const T &, Comp) [with IterT=stdext::checked_array_iterator<const vtkm::Int64 *>, T=vtkm::Id, Comp=vtkm::SortLess]" 

We have two options here:

1. Remove the checked iterator and just store the pointer on all platforms.
2. Only enable the checked iterator when compiling without nvcc.

I'm leaning towards 1. The implementation will be cleaner, and memcheck/cuda-memcheck are great at finding the sorts of issues the checked_array_iterator detects. IIRC, the warnings mentioned in the comment are disabled via MSVC's SECURE_NO_WARNINGS flags.

Option 2 seems to get into some weird binary compatibility issues, and prevents iterators from being able to be passed between host <-> device.

Does anyone else have input on this?