Vectorized (GPU) cfd solver

[RiccardoRossi]

📝 Description

this PR adds a semi explicit CFD solver which can either run on the CPU or GPU. Myself and @rubenzorrilla we have been working on this for quite a while and it is currently in a reasonable state.

it requires the library cupy and pyamg

python3 - m pip install cupy 
python3 -m pip install pyamg

The solver is entirely written in python but leverages blas calls to be efficient.
With the current version and using a mid-tier GPU it is about 10 times faster than the gpu.
Having said this, honestly the style is a good fit to the gpu but probably not for use on cpu. Efficient and parallel use of the CPU will depend on the undetying blas.

user can set gpu or OpenMP as:

    "problem_data": {
        "problem_name": "Drywell_CFD_Only",
        "parallel_type": "gpu", //"OpenMP"
	"precision" : "float32",
        "echo_level": 0,
        "start_time": 0.0,
        //"end_time": 0.02
        "end_time": 4.0
    },

[loumalouomega]

Chery pick this into an independent PR

[RiccardoRossi]

I woukd prefer to ñeave this together as it is needed for this pr.

[loumalouomega]

You can keep it here, but can be cherry-piced and merged in advance there, so it simplifies review here and keeps it simple

[rubenzorrilla]

I agree @loumalouomega. Indeed, I had this pending to be done. @RiccardoRossi I can do it.

[rubenzorrilla]

OK. I just saw that Charlie did it.

[loumalouomega]

And the test as well

[loumalouomega]

Minor comment related to changes that can be cherry-picked to another PR

[roigcarlo]

I've move the changes to #14413 but they will not work without part of the changes in the solver...

[loumalouomega]

I've move the changes to #14413 but they will not work without part of the changes in the solver...

Well, only what is actully standalone can be ported...Anyway, CI is failing...

[loumalouomega]

FYI: https://www.modular.com/open-source/mojo, is python GPU accelerated