Reflect padding output seems incorrect when padding size larger than input dimension

[ttyio]

Describe the bug
Reflect padding output seems incorrect when padding size larger than input dimension

Urgency
N/A

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Ubuntu 18.04
• ONNX Runtime installed from (source or binary): binary
• ONNX Runtime version: 1.8.0
• Python version: 3.6.9

To Reproduce
test_pad.zip

np.set_printoptions(threshold=sys.maxsize)
sess = rt.InferenceSession("test_pad.onnx")
input_name = sess.get_inputs()[0].name
X = np.arange(3*3).reshape(1, 1, 1, 3, 3).astype(np.float32)
pred_onx = sess.run(None, {input_name: X})[0]
print("onnxruntime:")
print(pred_onx)

Expected behavior
Not know the ground truth, but the output should at least symmetric, also should not a lot of 0s there.

Screenshots

[stale]

This issue has been automatically marked as stale due to inactivity and will be closed in 7 days if no further activity occurs. If further support is needed, please provide an update and/or more details.

[talc23]

System information

OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Ubuntu 20.04
ONNX Runtime installed from (source or binary): binary
ONNX Runtime version: 1.12.1
Python version: 3.8.10

can be reproduced using this script:
`
import onnxruntime as ort
import onnx
from onnx import helper
from onnx import TensorProto
import numpy as np

N = 4
padding = [N, 0]
X = helper.make_tensor_value_info('X', TensorProto.FLOAT, [N])
pads = helper.make_tensor('pads', TensorProto.INT64, vals=padding, dims=[2])
Y = helper.make_tensor_value_info('Y', TensorProto.FLOAT, [N+N])

node_def = helper.make_node(
'Pad', # name
['X', 'pads'], # inputs
['Y'], # outputs
mode='reflect', # attributes
)
graph_def = helper.make_graph(
[node_def], # nodes
'test-model', # name
[X],
[Y], # outputs
[pads]
)
model_def = helper.make_model(graph_def, producer_name='onnx-example')
onnx.checker.check_model(model_def)
print('The model is checked!')
onnx.save_model(model_def, 'pad.onnx')
print('The model is saved!')

x = np.arange(N, dtype=np.float32)
print(f'input={x}')
print(f'padding={padding}')
result = np.pad(x, pad_width=(padding[0], padding[1]), mode='reflect')
print(f'numpy result={result}')

y_ort = ort.InferenceSession("pad.onnx", providers=['CPUExecutionProvider']).run(["Y"], {"X": x})[0]
print(f'onnxrt result={y_ort}')
`

result should be:
The model is checked! The model is saved! input=[0. 1. 2. 3.] padding=[4, 0] numpy result=[2. 3. 2. 1. 0. 1. 2. 3.] onnxrt result=[0. 3. 2. 1. 0. 1. 2. 3.]