We always do one upcall_import to access the crate-cache, even though on a cache hit the upcall is not doing anything more complex than a couple indexed-load-and-compare checks and a return.

This is not necessary: we can emit those load/compare checks as a fast(er) hit path that stays in the stub, and only upcall on miss.

Deferring for now because it's still not a big cost at the moment, and there's more work to do around the stubs anyways in terms of refcounting.