Add user option for column-major rank assignment.

[romerojosh]

This PR introduces a new environment variable option, CUDECOMP_USE_COL_MAJOR_RANK_ORDER, to enable users to change the default row-major rank assignment over pdims to column-major.

This can be useful when users have problems that require process grids like 2 x 8 or 4 x 8 on a system with say, 4 NVLink connected GPUs per node (i.e. a small first dimension that is less than the number of NVLink connected GPUs). With the default row-major rank assignment, ranks are first assigned contiguously over the second dimension. This means that the first dimension, even though it is less than the number of NVLink connected GPUs, will span multiple nodes, resulting in slower internode communication for both the row and column communicators. Instead, if the rank assignment were column-major in this case, the ranks are first assigned across the small first dimension, resulting in a fast column communicator restricted to an NVLink connected group. This new option will enable users to realize this benefit in this type of situation.