Use datastack path for __init__ dispatch in CallAllocAndEnterInit (#8448)

specialization_run_init used prepare_exact_args_frame which heap-allocates
a FrameObject with 4-5 atomic refcount bumps per call. Switch to
invoke_prepared_exact_args which uses InterpreterFrame::new_on_datastack
with zero refcount bumps.

CallAllocAndEnterInit already guards against tracing (eval_frame_active)
and recursion limits before reaching this path.

Assisted-by: Claude

crates/vm/src/builtins/function.rs

@@ -791,7 +791,7 @@ impl Py<PyFunction> {
         frame
     }
 
-    fn invoke_prepared_exact_args(
+    pub(crate) fn invoke_prepared_exact_args(
         &self,
         args: impl ExactSizeIterator<Item = PyObjectRef>,
         vm: &VirtualMachine,

crates/vm/src/frame.rs

@@ -2673,10 +2673,7 @@ impl ExecutingFrame<'_> {
             .iter_mut()
             .map(|slot| slot.take().expect("arg slot must be filled"));
 
-        let init_frame = init_func.prepare_exact_args_frame(taken, vm);
-        let init_result = vm.run_frame(init_frame.clone());
-        release_datastack_frame(&init_frame, vm);
-        let init_result = init_result?;
+        let init_result = init_func.invoke_prepared_exact_args(taken, vm)?;
 
         if !vm.is_none(&init_result) {
             return Err(vm.new_type_error(format!(