Refactoring: replaced TNL::SharedPointer with std::shared_ptr for the linear solver and preconditioner

The DeviceType capability does not matter because these objects are
never passed to the device and we need to pass "shared_ptr<Derived>" as
"shared_ptr<Base>".

The vector view types have to be passed by value, because pass by
reference does not consider the converting constructors.

It is actually used only in 2 methods and unnecessarily complicates the
interface for all other solvers.

They are disabled and never really worked for these objects...

Works fine with g++ 5.5 and nvcc 8.0 in Arch Linux. File a bug report on
Gitlab if you have problems.

It is not needed, since CUSPARSE is included in all versions of CUDA
since at least 7.0.

Also note that CUSPARSE is needed only for some tests and legacy code
(tnl-benchmark-spmv and wrappers in Matrices::CSR) and experimental code
(Preconditioners::ILU0). Everything is still guarded by the
HAVE_CUSPARSE macro, so the user can pass -DHAVE_CUSPARSE -lcusparse to
the compiler to enable the features.

I don't understand how it could work on the small grids...

Preapare Distributed Grids for Communication groups (aka MPI Communicator)

Why do problems have this method anyway? It was obviously never used...

Added parallel reduction to communicators.