`#[plugin(...)]` at the top level should cause an error; `#![plugin(...)]` is correct

[apoelstra]

Right now if you write #[plugin(...)] to import some syntax extension, the import simply won't happen and other errors will occur. For example, if you write #[plugin(serde_macros)] and then later do #[derive(Serialize)] on a type you will get errors about Serialize not being implemented, as though the derive had done nothing.

This is extremely confusing since a user is basically being told "#derive had no effect" when the real problem is a missing exclamation mark in a totally different source file.

This should trigger an error to use #![plugin(...)] instead.

cc @erickt

[jonas-schievink]

The same works for no_std and probably all other gated attributes, since they are ignored by the unused_attributes lint. In other cases, the lint already prints "crate-level attribute should be an inner attribute: add an exclamation mark: #![foo]", but gated attributes cannot simultaneously be marked as CrateLevel.