Remove dead code and extend test coverage and diagnostics around it

We lost the following comment during refactorings:

The current code for niche-filling relies on variant indices instead of actual discriminants, so enums with explicit discriminants (RFC 2363) would misbehave.

crates/hir-ty/src/layout/adt.rs

@@ -3,9 +3,9 @@
 use std::{cmp, ops::Bound};
 
 use hir_def::{
-    AdtId, VariantId,
     layout::{Integer, ReprOptions, TargetDataLayout},
     signatures::{StructFlags, VariantFields},
+    AdtId, VariantId,
 };
 use intern::sym;
 use rustc_index::IndexVec;
@@ -13,9 +13,9 @@ use smallvec::SmallVec;
 use triomphe::Arc;
 
 use crate::{
-    Substitution, TraitEnvironment,
     db::HirDatabase,
-    layout::{Layout, LayoutError, field_ty},
+    layout::{field_ty, Layout, LayoutError},
+    Substitution, TraitEnvironment,
 };
 
 use super::LayoutCx;
@@ -85,16 +85,6 @@ pub fn layout_of_adt_query(
                 let d = db.const_eval_discriminant(e.enum_variants(db).variants[id.0].0).ok()?;
                 Some((id, d))
             }),
-            // FIXME: The current code for niche-filling relies on variant indices
-            // instead of actual discriminants, so enums with
-            // explicit discriminants (RFC #2363) would misbehave and we should disable
-            // niche optimization for them.
-            // The code that do it in rustc:
-            // repr.inhibit_enum_layout_opt() || def
-            //     .variants()
-            //     .iter_enumerated()
-            //     .any(|(i, v)| v.discr != ty::VariantDiscr::Relative(i.as_u32()))
-            repr.inhibit_enum_layout_opt(),
             !matches!(def, AdtId::EnumId(..))
                 && variants
                     .iter()