0.9.0

• Use NUMA-friendly internal data structures.
  This shows measurable speed-up on NUMA systems.
• Allow asynchronous amg setup.
  amgcl::amg constructor starts the setup process in a new thread. As soon
  as constructor returns, the instance is ready to be used as a
  preconditioner. Initially its just a single-level smoother, but when as
  the new (coarser) levels are constructed, they are put to use.
  In case of GPGPU backends, this should allow to overlap work between
  host CPU doing setup and the compute device doing the solution. In some
  cases a 2x speedup of the overall solution has been achieved.
• Allow limiting number of amg levels, thus supporting using relaxation
  for coarse solves.
• Rewrite lgmres and fgmres in terms of Givens rotations, which should
  work better with complex problems, see #34.
• Use new, more effective, sparse matrix format in VexCL backend and
  allow to use non-scalar values with the backend.
• Modernize cmake scripts.
  Provide amgcl::amgcl imported target, so that users may just

  find_package(amgcl)
  add_executable(myprogram myprogram.cpp)
  target_link_libraries(myprogram amgcl::amgcl)

  to build a program using amgcl. The imported target brings necessary
  compile and link options automatically.
• Replace boost.python with pybind11 and
  improve python interface.
• Unify example codes for different backends.
• Minor improvements and bug fixes