CUDA kernels should not ever work with distributed data structures, they
should always get the underlying *local* data structure.

- The "NullGroup" should not be used even when built without MPI,
  otherwise the behaviour is very bug-prone because "NullGroup" usage is
  not caught and changing the build type leads to a different semantics.
- "AllGroup" is not a good default value for the parameters, considering
  that the class attributes are initialized to "NullGroup".

Buffering with a small value is very slow when profiling anything in nvpp.

Amends e5fc6a96

Scan refactoring

Closes #87

See merge request !100

- structs from HorizontalOperations.h reimplemented as function objects in
  Functional.h
- repetitive function definitions generated using macros
- added new operators: % (modulus) and ^ (xor)

This way we test both the general CUDA implementation using shared
memory and the specialization using __shfl instructions.

Both the reduction and scan kernels needed some tweaks due to shared
memory usage with non-fundamental types.

This is needed because custom specializations of std::is_arithmetic
cannot be used (they cause an undefined behaviour).

The algorithms are implemented as plain functions in TNL::Algorithms.
containsValue was replaced with contains.

Removed reduction methods from Array and ArrayView, instead added overloads of reduce and reduceWithArgument for arrays/views

Plain functions are much more flexible than methods. The methods were
also violating the open-closed principle:
https://en.wikipedia.org/wiki/Open%E2%80%93closed_principle

This adds back the original approach (prescan + uniform shift) which
was removed too early.

The original approach (prescan + uniform shift) is more efficient for
inputs that are expensive to evaluate, such as vector expressions.