Revisit the request management login for the nonblocking resilience support

[bosilca]

Based on the discussions around #12384, it appears that we are not correctly handling the requests passed to the user in the ishrink case. After discussion with @abouteiller we have identified a proper way to handle this, and need to change and document the code to fit. Once done for ishrink, we should revisit all the nonblocking fault management functions, as well as the nonblocking communicator creation.