Handle Constant op variants in shape inference

[gramalingam]

Bug Report

Describe the bug

Shape-inference does not handle variants of Constant op introduced in opset 12. The opset 12 version introduced the ability to specify constant values using one of multiple attributes (like value_ints, for example) but the implementation handles only the original form (using the attribute value).

Reproduction instructions

• Describe the code to reproduce the behavior.

from onnx import shape_inference, parser

def infer(x):
    model = parser.parse_model(x)
    inferred_model = shape_inference.infer_shapes(model)
    return { vi.name : vi.type for vi in inferred_model.graph.value_info }

result = infer('''
<
  ir_version: 7,
  opset_import: [ "" : 15 ]
>
toy (int64[?] X) => (int64[] Y)
{
    S = Constant <value_ints = [10, 20]>()
    Z = Reshape (X, S)
    Y = Identity (Z)
}
''')

print(result['Z'])

result2 = infer('''
<
  ir_version: 7,
  opset_import: [ "" : 15 ]
>
toy (int64[?] X) => (int64[] Y)
{
    S = Constant <value = int64[2] {10, 20}>()
    Z = Reshape (X, S)
    Y = Identity (Z)
}
''')

print(result2['Z'])

Expected behavior

Both result and result2 should have same value

Notes

This line needs to be fixed:

onnx/onnx/shape_inference/implementation.cc

Line 341 in f0c4ee2

if (attr.name() == "value") {

[gramalingam]

This was completed by this PR: #4824