Enable execution of MG3 on GPU through OpenACC

[sjsprecious]

In this PR, we add the OpenACC directives to MG3 codes so that the MG3 parameterization could be run on GPU. We also make a few code changes for some calculations, which would lead to a non-BFB result.

The detailed changes related to OpenACC include:

• Add parallel directives to loops where the calculations are independent between columns and levels, and offload these calculations to GPU.
• Add data directives to explicitly manage the data movement between host and device.
• Add default(present) clause to ensure that data are on the device as expected before we perform a GPU computation.
• Add a tunable variable VLEN for the vector length.

The detailed changes other than OpenACC directives include:

• Change the avg_diameter elemental function to avg_diameter_vec subroutine.
• Move the calculation of gamma_bg_plus1 and gamma_bg_plus4 to the initialization subroutine and do it only once.
• Declare some constants as parameter in the pumas_util code.

Note that the change related to avg_diameter_vec subroutine is a non-BFB change for CPU run but passes the ensemble consistency test (ECT).

Test suite: [aux_cam] Test baseline: generated with CAM codes from https://github.com/PUMASDevelopment/CAM/tree/sjsp_cam_gpus_p3 and PUMAS codes from https://github.com/ESCOMP/PUMAS/tree/sjsp_cam_gpus_pumas_p3 Test machine: [Cheyenne] Test result: [NBFB]

[Katetc]

I'm going to merge this directly into the release/cam branch so I can make the tag for your CAM PR easier.

[sjsprecious]

I'm going to merge this directly into the release/cam branch so I can make the tag for your CAM PR easier.

Thanks Kate. That will be much helpful for our CAM PR!