Kokkos::sort support custom comparator

[fnrizzi]

Adds 2 overloads to Kokkos::sort to support a custom comparator.

template <class ExecutionSpace, class ComparatorType, class DataType, class... Properties>
void sort([[maybe_unused]] const ExecutionSpace& exec,
          const Kokkos::View<DataType, Properties...>& view,
          ComparatorType const& comparator)
{
   /* view must be rank-1 with layout{left,right,stride} */

  if constexpr (Impl::better_off_calling_std_sort_v<ExecutionSpace>){
    auto first = ::Kokkos::Experimental::begin(view);
    auto last  = ::Kokkos::Experimental::end(view);
    std::sort(first, last, comparator);
  } else {
    Impl::sort_device_view_with_comparator(exec, view, comparator);
  }
}

template <class ComparatorType, class DataType, class... Properties>
void sort(const Kokkos::View<DataType, Properties...>& view, ComparatorType const& comparator){
  // calls overload above using the view execution space
}

where:

Impl::sort_device_view_with_comparator(exec, view, comparator) 

calls cuda thrust for cuda, onedpl and the fallback case (for now) copies view to host, calls std::sort and copies back.
This solution is based on a discussion with @crtrott and @dalg24.

Note that this is potentially the place where we can later optimize for example implementing our own quicksort or what not.

[fnrizzi]

retest this please

[fnrizzi]

retest this please

[fnrizzi]

retest this please

[fnrizzi]

Only CI failure is related to #6293

[dalg24]

Would approve once we clarify whether the fence is needed or not

[dalg24]

The failure is unrelated (timeout during one of the HIP build)

[crtrott]

We could consider later to keep a buffer around (say 500k or so) which would for smaller views allow us to avoid the costly alloc/dealloc in the host copy case. But we can do that later.

[crtrott]

Failure looks unrelated.