Refactor MPI handling in TPS

[uvilla]

Interfaces with python/PARLA require that the MPI_Init is called from the main executable and not by the TPS::Tps class.

To this aims, this PR introduce the following changes:

• Remove the object mfem::MPI_Session from the base code. Note that mfem::MPI_Session is deprecated in mfem and does not allow for MPI initialization outside this object.

• Bump required MFEM version to 4.4

• Introduce an MPI_Comm TPSCommWorld to be used inside the tps code as the world communicator. New tests were added to run a simulation on a subset of processes

• Remove mfem::MPI_Session from the constructor of all TPS::Solver. TPS::Solver can grab the MPI_Comm TPSCommWorld directly from the TPS::Tps that is passed at construction time.

• All tps executables now call mfem::Mpi::Init from the main function

• tps.py now uses mpi4py to initialize MPI. MPI Communicator can now be provided from the Python interface

TODOs:

• [x] pip install mpi4py in the tps Docker container
• [x] Minimum MFEM version 4.4
• [x] Test the GPU built (Depends on PR #213 )
• [x] Update local runners to MFEM version 4.4 (Depends on PR #213 )
• [x] Allow the Python interface to wrap mpi4py communicators