Put the IR for the return type in the same block as the IR for the parameters. (#3145)

Previously, IR generated for the computation of the return type ended up
in the package's node block. Put it with the IR for the parameters for
consistency and cleanliness.

toolchain/semantics/semantics_handle_parameter_list.cpp

@@ -19,14 +19,6 @@ auto HandleDeducedParameterListStart(Context& context,
 auto HandleParameterList(Context& context, ParseTree::Node parse_node) -> bool {
   auto refs_id = context.ParamOrArgEnd(
       /*for_args=*/false, ParseNodeKind::ParameterListStart);
-  // TODO: This contains the IR block for parameters. At present, it's just
-  // loose, but it's not strictly required for parameter refs; we should either
-  // stop constructing it completely or, if it turns out to be needed, store it.
-  // Note, the underlying issue is that the LLVM IR has nowhere clear to emit,
-  // so changing storage would require addressing that problem. For comparison
-  // with function calls, the IR needs to be emitted prior to the call.
-  context.node_block_stack().Pop();
-
   context.PopScope();
   context.node_stack()
       .PopAndDiscardSoloParseNode<ParseNodeKind::ParameterListStart>();
@@ -44,7 +36,6 @@ auto HandleParameterListStart(Context& context, ParseTree::Node parse_node)
     -> bool {
   context.PushScope();
   context.node_stack().Push(parse_node);
-  context.node_block_stack().Push();
   context.ParamOrArgStart();
   return true;
 }