Add dependency to Rapids Memory Manager when building ParFlowGPU

[kvrigor]

@mfahdaz suggested that enabling Rapids Memory Manager (RMM) on a ParFlow CUDA build significantly boosts ParFlow performance. It might be best to enable RMM by default when building ParFlowGPU.

TODO:

• Add RMM to appropriate environment file/s.
• Add RMM_ROOT=/path/to/RMM/install to ParFlow build args

TSMP2/cmake/BuildParFlow.cmake

Lines 65 to 82 in 310a720

	ExternalProject_Add(ParFlow
	PREFIX ParFlow
	SOURCE_DIR ${PARFLOW_SRC}
	CMAKE_ARGS -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER}
	-DCMAKE_C_FLAGS=${PF_CFLAGS}
	-DCMAKE_CXX_COMPILER=${CMAKE_CXX_COMPILER}
	-DCMAKE_Fortran_COMPILER=${CMAKE_Fortran_COMPILER}
	-DCMAKE_Fortran_FLAGS=${PF_FFLAGS}
	-DCMAKE_BUILD_TYPE=${CMAKE_BUILD_TYPE}
	-DCMAKE_INSTALL_PREFIX=${CMAKE_INSTALL_PREFIX}
	-DPARFLOW_ENABLE_HYPRE=ON
	-DPARFLOW_ENABLE_NETCDF=ON
	-DPARFLOW_AMPS_SEQUENTIAL_IO=ON
	-DPARFLOW_ENABLE_TIMING=TRUE
	-DPARFLOW_ACCELERATOR_BACKEND=${PF_ACC_BACKEND}
	-DMPIEXEC_EXECUTABLE=${MPIEXEC_EXECUTABLE}
	-DMPIEXEC_NUMPROC_FLAG=${MPIEXEC_NUMPROC_FLAG}
	-DPARFLOW_ENABLE_SLURM=${ENABLE_SLURM}

[DCaviedesV]

Yes, this is a must. I thought we had this already, it was the case in the TSMP1 solution.