Replace a call to PyTuple_New with _PyTuple_FromArraySteal

PyTuple_New will zero out the tuple before returning to the caller, and a
surprising amount of time can be saved by not doing this zeroing.  One option
is to add a non-zeroing version of PyTuple_New, which I did in python#96446, but
there was resistance to the unsafety of it.

Fortunately it looks like most of the tuple-zeroing happens directly from the
BUILD_TUPLE opcode in the interpreter, which already has the arguments in an
appropriate array, so we can just convert this to _PyTuple_FromArraySteal

This seems to result in a ~0.2% speedup on macrobenchmarks.

Author: kmod

Python/ceval.c

@@ -2647,13 +2647,10 @@ _PyEval_EvalFrameDefault(PyThreadState *tstate, _PyInterpreterFrame *frame, int
         }
 
         TARGET(BUILD_TUPLE) {
-            PyObject *tup = PyTuple_New(oparg);
+            STACK_SHRINK(oparg);
+            PyObject *tup = _PyTuple_FromArraySteal(stack_pointer, oparg);
             if (tup == NULL)
                 goto error;
-            while (--oparg >= 0) {
-                PyObject *item = POP();
-                PyTuple_SET_ITEM(tup, oparg, item);
-            }
             PUSH(tup);
             DISPATCH();
         }