After further investigation, the real problem is not about how link options are specified by MPI module but rather how options are forwarded to the underlying compiler (in your case gcc).
To propagate the options, CMake use the prefix -Xcompiler. Unfortunately, the options contains options with prefix -Wl and the option -Xcompiler cannot be used in this case.
To handle -Wl prefix, the option -Xlinker must be used instead (i.e. -Wl,opt1,opt2 must be transformed in -Xlinker=opt,opt2). To support this, It is required to re-work how CMake generates link options for CUDA compiler.
Now, for MPI module, I think the correct approach is to update the module to wrap link options inside $<HOST_LINK:...> genex to avoid propagating them to device link step. But I am not an expert of CUDA environments.
So you can create an issue for MPI module describing the problem and add reference to this discussion. CUDA experts will be able to comment and validate (or not) my fix suggestion.
I will manage the creation of another issue for device link handling.