Before this PR, the distinction between "user code" and "system code" (e.g. core APIs) was based on the crate "doing the codegen", but #[inline]/generic fns (and even macros) can cause e.g. core internals to be codegen'd during the compilation of an user crate.

This used to not be a big issue, because "zombies" (i.e. deferred errors) only get silenced when they're in unused code, so they mostly served to hide errors in e.g. core internals that would not be codegen'd at all, had they had #[inline].

However, we've accumulated several features which rely on hiding various potentially-zombie things (such as the panic entry-point hijacking, or the "unused fn params" weakening), so we need to be using "zombies" (i.e. deferring errors) as much as possible, in order to actually take advantage of that infrastructure.

(I could've gone further and removed is_system_crate, forcing zombie_even_in_user_code's behavior always, but that might lead to confusing outcomes, so I decided against it - for now, anyway)