Revert "Optimize multi_tensor_apply (take 2) (#119764)"

This reverts commit 0b68a28c87df2c6eb2cf530be4659b5a2f8a95b0.

Reverted https://github.com/pytorch/pytorch/pull/119764 on behalf of https://github.com/huydhn due to Sorry for reverting your change but it is failing ROCm job in trunk 0b68a28c87.  Please help take a look and reland the change ([comment](https://github.com/pytorch/pytorch/pull/119764#issuecomment-2014190124))

test/test_foreach.py

@@ -649,26 +649,14 @@ class TestForeach(TestCase):
 
     @onlyCUDA
     @ops(foreach_reduce_op_db, allowed_dtypes=floating_types())
-    @parametrize("use_cuda_graph", (False, True))
-    def test_big_num_tensors(self, device, dtype, op, use_cuda_graph):
+    def test_big_num_tensors(self, device, dtype, op):
         N = 600
         tensorlist = [make_tensor((2, 3), dtype=dtype, device=device, noncontiguous=False) for _ in range(N)]
         fn, ref_fn, *_ = self._get_funcs(op)
 
         import math
         for ord in (1, 2, math.inf):
-            if not use_cuda_graph:
-                actual = fn(inputs=[tensorlist], is_cuda=True, expect_fastpath=True, ord=ord, zero_size=False)
-            else:
-                # When using CUDA graphs and the tensor metadata doesn't fit in
-                # the static kernel argument space, multi_tensor_apply creates
-                # the launch arguments once, uses cudaUserObject_t to tie its
-                # lifetime to the graph, and reuses it throughout replays. This
-                # test verifies multi_tensor_apply's behavior in the scenario.
-                g = torch.cuda.CUDAGraph()
-                with torch.cuda.graph(g):
-                    actual = fn.func(tensorlist, ord=ord)
-                g.replay()
+            actual = fn(inputs=[tensorlist], is_cuda=True, expect_fastpath=True, ord=ord, zero_size=False)
             expect = ref_fn(inputs=[tensorlist], ord=ord)
 
             self.assertEqual(expect, actual, equal_nan=True)