@llvm.assume blocks optimization

We don’t have that, AFAIK, but that is a good idea. Especially if we manifested equivalent information in the IR elsewhere we can drop assumptions for sure.