[feature request]Allow model developer to control launching method for pre_processing_stage, execution_stage and post_processing_stage

[zhiyuli]

Currently pre_processing_stage and post_processing_stage are launched in single process/threaded mode, execution_stage is ran in mpi mode through "srun --unbuffered --mpi=pmi2".

In some cases model developer would require more flexibility to control how each step is launched.
We could allow developer to specify a string/command in the manifest.json for each step and insert the string/command between and .

This will be a optional switch, and only override old sbatch file if they are present

for example, this following manifest will allow all the 3 steps to be launched by MPI

{
"name": "MPI Hello World",
"container": "mpich",
"pre_processing_stage": "python setup.py",

"pre_processing_stage_launching_command": "xxxxxxxxxx1",

"execution_stage": "python main.py",

"execution_stage_launching_command": "xxxxxxxxxxxxxxx2",

"post_processing_stage": "python cleanup.py",

"post_processing_stage_launching_command": "xxxxxxxxx3",

"slurm_ceiling": {
    "walltime": "2:00"
}

}