Use Stan's own MPI communicator to avoid communicator interference

Description

This is a following up, or more specific issue spawned from https://github.com/stan-dev/math/issues/939. Namely, currently Stan's MPI implementation uses a predefined global communicator MPI_COMM_WORLD, this is against the recommended use of the communicators, see

https://www.mcs.anl.gov/research/projects/mpi/tutorial/gropp/node116.html
https://www.open-mpi.org/doc/current/man3/MPI_Comm_dup.3.php#toc8

Specifically, any communication, even if using a local communicator(as an application should be), will be polluted by map_rect 's communicator, whether that local communicator is within or out of Stan. This causes unpredictable behavior, therefore is a design flaw.

Proposed short-term solution

Replace construction boost::mpi::communicator world calls with
boost::mpi::communicator world(MPI_COMM_WORLD, comm_duplicate)

Proposed long-term solution

Address https://github.com/stan-dev/math/issues/939.

Expected Output

With current tests, same. But need add more tests with new communicators.

Current Version:

v2.18.1

Locate the Boost.MPI communicator constructions used by Stan's MPI implementation, especially the world instances described in the issue, and review the existing MPI tests. Verify the communicator handling against the current tests, then add coverage using separate communicators and confirm that communication remains isolated.