fix:fix issue #238

[hush-coder]

fix: May I ask if I can fix issue #238

Current job:

• add infer_op_sets (Abs/Add/Relu/Cast/Concat/Compare/Conv1d/2d/3d/BatchNorm/MaxPool/AvgPool)
• add all_op_sets for CLI choices/validation; keep list(dict.keys()) style
• add --kind {train|inference}; require when op exists in both sets
• return JSON includes "type"; keep input_shapes as string for backward compatibility

Future work

• Implement '--json_file/--case/--case_id' in 'python/test/test_onnx.py' and 'python/test/test_torch.py';
• set up and test_fx. Py similar points, single operator to cover basic operator (Conv/Pool/Concat/Cast/Compare/BatchNorm/Relu/Abs/Add).
• extension infer_op_sets coverage: increase MatMul/Gemm/LeakyRelu/Sigmoid Tanh/Softmax, etc.;
• Perform more rigorous shape and symmetry checks on attributes such as' pads/dilations/axis', maintaining constraints consistent with the current training consumer end.

May I ask if this job is still necessary and whether my direction is correct?

[xujm1998]

您好，感谢您的工作；
test_fx.py主要是用于训练图的算子测试，这部分功能看起来没问题了。

但#238
主要还是想拓展test_onnx.py和test_torch.py的功能；
如果时间有限可以先优先支持test_onnx.py。

onnx所支持的算子可以参考开发手册或python/transform/OnnxConverter.py