Licenses get to set terms of redistribution. But training of the ML model -- the thing described by your #3 -- is not redistribution (imho). So maybe it's as unenforceable as saying left-handed people can't read your code.

The redistribution happens later, either when copilot blurps out some of your code, or when the copilot user then distributes something using that code (I'm curious which). At that point, whether some use of your code is infringing your license doesn't depend on the path the code took, does it? (in which case #3 is moot)