Support user-defined types in radix sort and introduce initial sphinx docs

[gevtushenko]

This PR introduces support for user-defined types in the block / device radix sort (addresses #52):

struct custom_t
{
  std::uint16_t i;
  int unused;
  float f;
};

struct decomposer_t
{
  __host__ __device__ //
  ::cuda::std::tuple<std::uint16_t&, float&> operator()(custom_t &key) const
  {
    return {key.i, key.f};
  }
};

cub::DeviceRadixSort::SortKeys(d_temp_storage,
                               temp_storage_bytes,
                               d_in,
                               d_out,
                               num_items,
                               decomposer_t{});

There's no difference in performance between sorting a key that's a pair of int32 and sorting a single int64 key. The PR also doesn't affect SASS for fundamental types.

Documentation for the new API is written using sphinx, whose initial support is also introduced by this PR. Doxygen is temporarily broken, and a follow-up PR will provide the complete transition to sphinx. To build sphinx docs:

:cd docs 
:./gen_docs.sh 
:open _build/docs/CUB/latest/index.html

[elstehle]

Really solid work! Like how we can keep most of the implementation details unchanged.

Partial review; review of BlockRadixSort and tests remaining. Also want to push radix_rank_sort_operations through its paces a bit.

[elstehle]

Only a few, mostly minor, comments remaining 🙂