Improve invalid operator assignment handling.

[rcatolino]

So far the following code

struct Foo;

fn main() {
  let mut t = Foo;
  let ref b = Foo;
  a += *b;
}

errors with

test.rs:15:3: 13:11 error: binary operation + cannot be applied to type `Foo`
test.rs:15   *a += *b;

Since assignment-operators are no longer expanded to left = left OP right but are independents operators it should be

test.rs:15:3: 13:11 error: binary operation += cannot be applied to type `Foo`
test.rs:15   *a += *b;

to make it clear that implementing Add for Foo is not gonna work. (cf issues #11143, #11344)

Besides that, we also need to typecheck the rhs expression even if the operator has no implementation, or we end up with unknown types for the nodes of the rhs and an ICE later on while resolving types. (once again cf #11143 and #11344).

This probably would get fixed with #5992, but in the meantime it's a confusing error to stumble upon.
@pcwalton, you wrote the original code, what do you think?
(closes #11143 and #11344)

[alexcrichton]

It'd be nice to have a test for the new functionality in the compile-fail suite of tests.

The change to the enums and error messages seem fine to me, but what's with the new check_expr(fcx, rhs);, that's just recursing to perform more typechecking, right?

[rcatolino]

Yes the check_expr is there because of the typechecking, issue I mentioned. (Causes the ICE described in #11143 and #11344).
Test added. Just to be sure, the tests in compile-fail will fail if an ICE arises right? Otherwise this one is quite useless.
About the enum, I changed it because semantically the old name wouldn't have been correct anymore : we are now using it to check the context of the BinOp (assignment or not) and not just whether we allow overload. I didn't want to confuse the people working on it in the future.

[rcatolino]

oops, a test in compile-fail wasn't expecting this error message, I updated it. Retry?