Merge PR MPAS-Dev#26: Implement OpenACC-friendly halo exchanges

This merge introduces new infrastructure for OpenACC-friendly halo exchanges for
2-d and 3-d real fields, and it updates the halo exchange calls in the MPAS-Atmosphere
dynamics to use new routines for exchanging halos on both CPUs and GPUs.

Existing halo exchange routines that work on MPAS field types require the use of
complicated "multi-halo exchange lists" within the halo exchange routines themselves.
Transferring these data structures onto GPUs to allow direct GPU-to-GPU halo
exchanges is overly complicated with OpenACC, so new routines (collectively using
the interface `mpas_dmpar_exch_halo_acc`) have been developed.

The `mpas_dmpar_exch_halo_acc` routines make use of a simple array of integers to
describe the communication of elements with other MPI tasks, and are thus are more
easily mapped from host to GPU memory.

The example, below, illustrates the use of the new OpenACC-friendly routines as they
are applied in the MPAS-Atmosphere dynamics:

```
  integer, dimension(:), pointer :: compactHaloInfo_c_v
  type (field2DReal), pointer :: field_2d_ptr
  real (kind=RKIND), dimension(:,:), pointer :: theta_m

  ! Build integer array of halo information for a given field dimensionality
  call mpas_pool_get_field(state, 'rho_zz', field_2d_ptr, 1)
  call mpas_dmpar_compact_halo_info(field_2d_ptr, compactHaloInfo_c_v)

  ! Exchange the halo for a field array using the halo information
  call mpas_pool_get_array(state, 'theta_m', theta_m, 1)
  call mpas_dmpar_exch_halo_acc(theta_m, compactHaloInfo_c_v)

  ! Halo information can be re-used as many times as needed before being deallocated
  deallocate(compactHaloInfo_c_v)
```

Note that the `mpas_pool_get_array_gpu` routine contains an `!ACC ENTER DATA COPYIN`
directive internally to check if data are present on the GPU or should be copied.
This facilitates the GPU direct communication.

Merge branch 'gpu_halos_merge' into atmosphere/gpu

* gpu_halos_merge: (24 commits)
  Replacing CPU halo exchange calls with CPU + GPU halo exachage calls.
  Initialize commListPtr = 0 when building send/recv lists for OpenACC exchanges
  Compute compactHaloInfo arrays needed by scalar transport in mpas_atm_dynamics_init
  Move computation of compactHaloInfo_* arrays into mpas_atm_dynamics_init
  Add empty mpas_atm_dynamics_{init,finalize} routines
  Pre-process out log writes in mpas_dmpar_exch_halo_acc if running on GPUs
  Convert all halo exchanges in atm_srk3() to mpas_dmpar_exch_halo_acc interface
  Remove a few debugging write statements from mpas_dmpar_exch_halo_acc routines
  Only use compactHaloInfo and field arrays in mpas_dmpar_exch_halo_acc routines
  Build {send,recv}CommList exclusively from compactHaloInfo in mpas_dmpar_build_comm_lists_acc
  Add endpoint IDs to compactHaloInfo array
  Use compactHaloInfo to obtain product of inner dims in mpas_dmpar_build_comm_lists_acc
  Add sendList and recvList information to compactHaloInfo array
  Use constant layout of compactHaloInfo for all field types
  Add mpas_dmpar_build_comm_lists_acc routine to create comm lists from compactHaloInfo
  In mpas_dmpar_exch_halo_acc routines, get nHaloLayers from compactHaloInfo
  Add 'array' argument to mpas_dmpar_exch_halo_acc routines
  Get MPI communicator and rank info from compactHaloInfo in mpas_dmpar_exch_halo_acc
  mpas_dmpar_exch_halo_acc routines now require compactHaloInfo argument
  Add interface mpas_dmpar_compact_halo_info
  ...

Author: mgduda