Rollup merge of #110930 - b-naber:normalize-elaborate-drops, r=cjgillot

Don't expect normalization to succeed in elaborate_drops

Fixes https://github.com/rust-lang/rust/issues/110682

This was exposed through the changes in https://github.com/rust-lang/rust/pull/109247, which causes more things to be inlined. Inlining can happen before monomorphization, so we can't expect normalization to succeed. In the elaborate_drops analysis we currently have [this call](033aa092ab/compiler/rustc_mir_dataflow/src/elaborate_drops.rs (L278)) to `normalize_erasing_regions`, which ICEs when normalization fails. The types are used to infer [whether the type needs a drop](033aa092ab/compiler/rustc_mir_dataflow/src/elaborate_drops.rs (L374)), where `needs_drop` itself [uses `try_normalize_erasing_regions`](033aa092ab/compiler/rustc_middle/src/ty/util.rs (L1121)).

~[`instance_mir`](https://doc.rust-lang.org/stable/nightly-rustc/rustc_middle/ty/context/struct.TyCtxt.html#method.instance_mir) isn't explicit about whether it expects the instances corresponding to the `InstanceDef`s to be monomorphized (though I think in all other contexts the function is used post-monomorphization), so the use of `instance_mir` in inlining doesn't necessarily seem wrong to me.~

compiler/rustc_mir_dataflow/src/elaborate_drops.rs

@@ -276,6 +276,7 @@ where
                 assert_eq!(self.elaborator.param_env().reveal(), Reveal::All);
                 let field_ty =
                     tcx.normalize_erasing_regions(self.elaborator.param_env(), f.ty(tcx, substs));
+
                 (tcx.mk_place_field(base_place, field, field_ty), subpath)
             })
             .collect()