coercion ?0: Sized check, rewrite to not rely on fulfill internals

[lcnr]

The behavior of fn coerce_unsized for some inference variable ?0 changes depending on whether there's a pending ?0: Sized obligation in the fulfillment context:

rust/compiler/rustc_hir_typeck/src/coercion.rs

Lines 672 to 682 in 63c748e

	match selcx.select(&obligation.with(trait_pred)) {
	// Uncertain or unimplemented.
	Ok(None) => {
	if trait_pred.def_id() == unsize_did {
	let trait_pred = self.resolve_vars_if_possible(trait_pred);
	let self_ty = trait_pred.skip_binder().self_ty();
	let unsize_ty = trait_pred.skip_binder().trait_ref.substs[1].expect_ty();
	debug!("coerce_unsized: ambiguous unsize case for {:?}", trait_pred);
	match (&self_ty.kind(), &unsize_ty.kind()) {
	(ty::Infer(ty::TyVar(v)), ty::Dynamic(..))
	if self.type_var_is_sized(*v) =>

This pending obligation can be a nested obligation, e.g. we may have only added a Vec<?0>: Debug obligation to the fulfillment context. But proving that adds a nested ?0: Sized obligation from the impl candidate to the FulfillmentContext, which then influences coercion.

As our new solver should try to fully prove obligations instead of selecting a candidate and adding nested obligations to the root, this has to change.

cc @rust-lang/initiative-trait-system-refactor

[compiler-errors]

#50753 is related