Shrink SelectionError a lot

`SelectionError` used to be 80 bytes (on 64 bit). That's quite big.
Especially because the selection cache contained `Result<_,
SelectionError>. The Ok type is only 32 bytes, so the 80 bytes
significantly inflate the size of the cache.

Most variants of the `SelectionError` seem to be hard errors, only
`Unimplemented` shows up in practice (for cranelift-codegen, it occupies
23.4% of all cache entries). We can just box away the biggest variant,
`OutputTypeParameterMismatch`, to get the size down to 16 bytes, well
within the size of the Ok type inside the cache.

compiler/rustc_middle/src/traits/mod.rs

@@ -580,11 +580,7 @@ pub enum SelectionError<'tcx> {
     /// After a closure impl has selected, its "outputs" were evaluated
     /// (which for closures includes the "input" type params) and they
     /// didn't resolve. See `confirm_poly_trait_refs` for more.
-    OutputTypeParameterMismatch(
-        ty::PolyTraitRef<'tcx>,
-        ty::PolyTraitRef<'tcx>,
-        ty::error::TypeError<'tcx>,
-    ),
+    OutputTypeParameterMismatch(Box<SelectionOutputTypeParameterMismatch<'tcx>>),
     /// The trait pointed by `DefId` is not object safe.
     TraitNotObjectSafe(DefId),
     /// A given constant couldn't be evaluated.
@@ -596,6 +592,13 @@ pub enum SelectionError<'tcx> {
     ErrorReporting,
 }
 
+#[derive(Clone, Debug, TypeVisitable, Lift)]
+pub struct SelectionOutputTypeParameterMismatch<'tcx> {
+    pub found_trait_ref: ty::PolyTraitRef<'tcx>,
+    pub expected_trait_ref: ty::PolyTraitRef<'tcx>,
+    pub terr: ty::error::TypeError<'tcx>,
+}
+
 /// When performing resolution, it is typically the case that there
 /// can be one of three outcomes:
 ///