- LB on GPU, Uniform Grid Benchmark app
- helper class to schedule tasks to multiple CUDA streams

previously the ghost layer was aligned

Features:
   - uses generated pack infos for packing & unpacking directly on GPU
   - can directly send GPU buffers if cuda-enabled MPI is available,
     otherwise the packed buffers are transfered to CPU first
   - communication hiding with cuda streams: communication can be run
     asynchronously - especially useful when compute kernel is also
     split up into inner and outer part

- added RAII classes for CUDA streams and events
- equivalence test that checks if generated CPU and GPU (overlapped)
  versions are computing same result as normal waLBerla LBM kernel

Part of issue #48

Changes introduced in this commit are the following:

- CUDA streams: Add support for asynchronous (un)packing operations using CUDA
  streams in cuda::communication::GPUPackInfo. Through asynchronous operations
  it is possible to overlap GPU computation and MPI communication in simulations
  (e.g. LBM simulations). Asynchronous copies in CUDA require pinned memory on
  the host, and for that purpose a staging buffer is introduced (i.e.
  cuda::communication::PinnedMemoryBuffer) in the cuda module, which is used to
  stage data between the GPU and the MPI buffers.

- zyxf layout: Add zyxf field layout support in GPUPackInfo through extensions
  of the functions in cuda::GPUCopy.

- Extended GPUPackInfo test: Add stream and zyxf layout tests to the
  GPUPackInfoTest to test the proposed implementation.

- Extended Kernel: add CUDA stream and shared memory configuration support in
  cuda::Kernel class.

Signed-off-by: João Victor Tozatti Risso <joaovictortr@protonmail.com>