The issue is that the compiler we ship, GNU Fortran, assumes that Coarrays, regardless of the implementation, will be using multiple images.  This decision isn't arbitrary, though.  Many of the assumptions that the compiler itself makes assumes that the program is running in a self-contained image.  For example, when a variable has the SAVE attribute, meaning its value will persist across calls to a procedure, Coarray Fortran assumes that the saved value is local to that Coarray worker.  Yet the compiler treats it as a special static variable local to that process.  If Coarrays relied on threads, every thread, in theory, would see the same value in that variable because it is local to that process, not thread.  Those odd, little corners of the language makes implementing Coarrays without multiple processes highly problematic.

Launching this coarray program from a DLL would be exceptionally complicated because it needs to actually start additional processes for the Fortran coarray handling.  I'm not exactly sure how this situation would be handled from a call originating from a C++ of C# program because it would skip the multiple-image-initialization step. 

I think it might be more useful to look at what you wish to achieve.  You could have your C++ or C# program launch a Fortran executable that uses coarrays and read results from a file possibly.  If you're just looking to perform some parallel computations, OpenMP programming might be somewhat more convenient since it relies only on threads, which wouldn't run into the multiple-image issues I described.

Can you tell us more about what you're trying to achieve?  Coarrays aren't necessarily the best answer for parallel computations.

Thanks in advanced!