[fix]: fix tanh grad grad, add the third derivative test

[lucky9-cyou]

修复之前更新 tanh op 的高阶微分规则。使用tanh_grad来计算tanh_grad_grad中grad_require_grad的部分。添加了三阶微分的测试，二阶微分的测试不会涉及到grad_require_grad这一部分。

在OneScience中的cylinder2d例子的性能变化：
origin: 0.29/epoch
now: 0.26/epoch

[lucky9-cyou]

加的两个test会导致abs和square算子出现不一样的问题：

❯ python test_math_op_higher_derivative.py       
E.......................F..
======================================================================
ERROR: test_abs_grad_grad (__main__.TestMathOpHigherDerivative)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/data/home/zhangyanglin/Code/oneflow-master/python/oneflow/test_utils/automated_test_util/torch_flow_dual_object.py", line 668, in get_pytorch_oneflow_res
    pytorch_res = pytorch(*pytorch_args, **pytorch_kwargs)
  File "/data/home/zhangyanglin/miniconda3/envs/oneflow-compile/lib/python3.9/site-packages/torch/autograd/__init__.py", line 300, in grad
    return Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "/data/home/zhangyanglin/Code/oneflow-master/python/oneflow/test/modules/test_math_op_higher_derivative.py", line 156, in test_abs_grad_grad
    _test_math_op_grad_grad_impl(test_case, "abs")
  File "/data/home/zhangyanglin/Code/oneflow-master/python/oneflow/test/modules/test_math_op_higher_derivative.py", line 52, in _test_math_op_grad_grad_impl
    x_grad_grad_grad = torch.autograd.grad(x_grad_grad, x, init_grad, retain_graph=True)[0]
  File "/data/home/zhangyanglin/Code/oneflow-master/python/oneflow/test_utils/automated_test_util/torch_flow_dual_object.py", line 830, in dual_method
    pytorch_res, oneflow_res = get_pytorch_oneflow_res(
  File "/data/home/zhangyanglin/Code/oneflow-master/python/oneflow/test_utils/automated_test_util/torch_flow_dual_object.py", line 733, in get_pytorch_oneflow_res
    raise PyTorchDoesNotSupportError(e)
oneflow.test_utils.automated_test_util.torch_flow_dual_object.PyTorchDoesNotSupportError: PyTorch error: element 0 of tensors does not require grad and does not have a grad_fn

======================================================================
FAIL: test_square_grad_grad (__main__.TestMathOpHigherDerivative)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/data/home/zhangyanglin/Code/oneflow-master/python/oneflow/test/modules/test_math_op_higher_derivative.py", line 150, in test_square_grad_grad
    _test_math_op_grad_grad_impl(test_case, "square")
  File "/data/home/zhangyanglin/Code/oneflow-master/python/oneflow/test/modules/test_math_op_higher_derivative.py", line 53, in _test_math_op_grad_grad_impl
    test_case.assertTrue(
AssertionError: False is not true

❯ python test_global_math_op_higher_derivative.py
F.......................F..
======================================================================
FAIL: test_global_abs_grad_grad (__main__.TestGlobalMathOpHigherDerivative)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/data/home/zhangyanglin/Code/oneflow-master/python/oneflow/test_utils/automated_test_util/torch_flow_dual_object.py", line 1370, in new_f
    return f(*args, **kwargs)
  File "/data/home/zhangyanglin/Code/oneflow-master/python/oneflow/test/modules/test_global_math_op_higher_derivative.py", line 242, in test_global_abs_grad_grad
    _global_math_op_grad_grad_impl(test_case, "abs", placement, sbp)
  File "/data/home/zhangyanglin/Code/oneflow-master/python/oneflow/test/modules/test_global_math_op_higher_derivative.py", line 45, in _global_math_op_grad_grad_impl
    test_case.assertTrue(
AssertionError: False is not true

======================================================================
FAIL: test_global_square_grad_grad (__main__.TestGlobalMathOpHigherDerivative)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/data/home/zhangyanglin/Code/oneflow-master/python/oneflow/test_utils/automated_test_util/torch_flow_dual_object.py", line 1370, in new_f
    return f(*args, **kwargs)
  File "/data/home/zhangyanglin/Code/oneflow-master/python/oneflow/test/modules/test_global_math_op_higher_derivative.py", line 230, in test_global_square_grad_grad
    _global_math_op_grad_grad_impl(test_case, "square", placement, sbp)
  File "/data/home/zhangyanglin/Code/oneflow-master/python/oneflow/test/modules/test_global_math_op_higher_derivative.py", line 56, in _global_math_op_grad_grad_impl
    test_case.assertTrue(
AssertionError: False is not true

[lucky9-cyou]

在两个文件的三阶导测试中，oneflow的square算子三阶导会跑出来nan，torch的square算子三阶导全是0

[lucky9-cyou]

在非global测试中，pytorch的abs算子三阶导会报上面的错误，global的测试中，pytorch的abs二阶导是错的，不是0，oneflow的是0，三阶导会报上面的错误。

[lucky9-cyou]

abs和square的bug放在了这个issue中：#10243

[levi131]

LGTM

[hjchen2]

在两个文件的三阶导测试中，oneflow的square算子三阶导会跑出来nan，torch的square算子三阶导全是0

torch的square算子三阶导我这里是None，你用的pytorch版本是多少