Fix a comment.

Imagine you have replace ranges (2..20,X) and (5..15,Y), and these tokens:
```
a,b,c,d,e,f,g,h,i,j,k,l,m,n,o,p,q,r,s,t,u,v,w,x
```
If we replace (5..15,Y) first, then (2..20,X) we get this sequence
```
a,b,c,d,e,Y,_,_,_,_,_,_,_,_,_,p,q,r,s,t,u,v,w,x
a,b,X,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,_,u,v,w,x
```
which is what we want.

If we do it in the other order, we get this:
```
a,b,X,_,_,_,_,_,_,_,_,_,_,_,_,p,q,r,s,t,u,v,w,x
a,b,X,_,_,Y,_,_,_,_,_,_,_,_,_,_,_,_,_,_,u,v,w,x
```
which is wrong. So it's true that we need the `.rev()` but the comment
is wrong about why.

Author: nnethercote

compiler/rustc_parse/src/parser/attr_wrapper.rs

@@ -137,9 +137,9 @@ impl ToAttrTokenStream for LazyAttrTokenStreamImpl {
             // `#[cfg(FALSE)] struct Foo { #[cfg(FALSE)] field: bool }`
             //
             // By starting processing from the replace range with the greatest
-            // start position, we ensure that any replace range which encloses
-            // another replace range will capture the *replaced* tokens for the inner
-            // range, not the original tokens.
+            // start position, we ensure that any (outer) replace range which
+            // encloses another (inner) replace range will fully overwrite the
+            // inner range's replacement.
             for (range, target) in replace_ranges.into_iter().rev() {
                 assert!(!range.is_empty(), "Cannot replace an empty range: {range:?}");