I don’t understand the problem. Adding a few lines of python to a script that rarely changes should be trivial in comparison to all the other changes necessary in a fork that adds a new backend. The overhead should be negligible, git pull will almost always just work, and when it doesn’t the conflict resolution should not be difficult. There’s very little churn in those scripts.
That is, option 2 is the intended way for downstreams to approach this, like extending anything else in LLVM, and I do not understand your justification for the added complexity you are proposing.