Generate macros for loops over any combination of dimensions, which get
pre-generated ranges so that each type of loop is parallelized over the
shared MPI arrays with optimal load balance.

Only functional difference between the different 'regions' is in
debugging code, but regions provide structure showing where
synchronization calls are needed.

Now replaced by the new implementation in looping.jl.

See JuliaParallel/MPI.jl#518.

If we are precompiling, presumably want an optimized, production build
so always pass `-O3 --check-bounds=no` to the build process.

Parsing command line options in command_line_options.__init__() caused
problems (command line arguments being ignored) when moment_kinetics was
compiled into a static system image. [My guess is that __init__() was
called too early in that case, before ARGS was set up properly.]

Instead, define a getter function get_options(), which parses the
command line arguments when it is called. This is slightly less
efficient, but will only be called a few times, and should be more
robust. Also allows changing the options, e.g. adding/removing "--long"
from ARGS during a REPL session to change which tests are run.

Also test with 3 processes, as this is now supported by moment_kinetics.