bind the value of tail expr before Ok-wrarpping

[ubnt-intrepid]

It changes the desugaring rule to bind the return value to a temporary variable before wrapping, rather than modifying the AST of tail expression.

This PR is a successor of #43 and the expansion rule comes originally from #43 (comment).

It fixes #42 at the same time since the wrapping expression always suppresses the warning message by #[allow(unreachable_code)].

[withoutboats]

Haha and honestly the code is much cleaner than what it was doing before! I'll test it out on some codebases locally to check that it works well and get back to you. Thanks!

[withoutboats]

So, this is a breaking change. But it's one that probably will strictly improve things, so I'm planning that we merge it and release a 2.0 at some point.

I found two breakages in my code that result from this.

The first is that the current code will allow some code to compile that really it shouldn't. In particular, consider this:

#[throws(())]
fn foo() {
    0
}

This code should not compile, because the final expression doesn't have the same type as the return type. Unfortunately, on the 1.0.0 branch it does, because we just insert a Ok(()) after it. In your branch, it correctly does not compile.

The second error I found arose from using throws in a trait definition. Now, I believe this causes the body to be expanded to let _ret = { ; }; .... So if the trait method was returning something, now you get a type error, when really you shouldn't. This is connected to the problem that #40 is aiming at: we don't really handle trait methods very well, and there are bugs around that. We should not be modifying the bodies of trait methods without bodies at all.

I'm going to merge this now, and then hopefully through #40 and maybe some subsequent PRs if necessary we can get it handling trait methods properly.