Modularised MPI training so methods can be resused for HPO. MPI training works fine. HPO is work in progress

Introduced custom data types to allows sending rows of a Phi matrix which is stored in a column-major order.

Work begins on redistribution of data. From 1d block cyclic to 2d block cyclic. Added routine to copy redistribute matrix, so far it just a simple copy.

Implemented OLD with scalapack. It has not been tested. Some work is requires to optimise data distribution across processes. For now it is always Nx1 grid

Almost there, just need to introduce work balancing scheme for cases where there is no more work and the worker's phi is not full