WIP: Add GPU aware MPI support in cannon algorithm #1308

Job	Run time
build-on-cuda (OPENMP=ON)	12m 5s
build-on-rocm (OPENMP=ON)	5m 4s
build-and-test (MPI=ON, OPENMP=ON, SMM=blas, openmpi)	2m 2s
build-and-test (MPI=ON, OPENMP=ON, SMM=blas, mpich)	2m 3s
build-and-test (MPI=ON, OPENMP=ON, SMM=libxsmm, openmpi)	2m 4s
build-and-test (MPI=ON, OPENMP=ON, SMM=libxsmm, mpich)	2m 2s
build-and-test (MPI=ON, OPENMP=OFF, SMM=blas, openmpi)	2m 3s
build-and-test (MPI=ON, OPENMP=OFF, SMM=blas, mpich)	1m 58s
build-and-test (MPI=ON, OPENMP=OFF, SMM=libxsmm, openmpi)	2m 6s
build-and-test (MPI=ON, OPENMP=OFF, SMM=libxsmm, mpich)	2m 7s
build-and-test (MPI=OFF, OPENMP=ON, SMM=blas, openmpi)	2m 4s
build-and-test (MPI=OFF, OPENMP=ON, SMM=libxsmm, openmpi)	2m 4s
build-and-test (MPI=OFF, OPENMP=OFF, SMM=blas, openmpi)	2m 4s
build-and-test (MPI=OFF, OPENMP=OFF, SMM=libxsmm, openmpi)	2m 3s
build-on-opencl (OPENMP=ON, SMM=libxsmm)	16m 59s
pre-commit	1m 27s
Combine & check coverage.	0s
	1h 0m 15s

Provide feedback