Fix issues with inference of onnx models converted from paddlex

[0x3878f]

1. PicoDet类模型

   In Node, ("ReduceMin.0", ReduceMin, "", -1) : ("Concat.17": tensor(int64),"helper.constant.25": tensor(int64),) -> ("ReduceMin.1",) , Error Node (ReduceMin.0) has input size 2 not in range [min=1, max=1]. 

   multiclass_nms3中ReduceMin在opset version 14~17时输入不正确

2. SOLOv2

   RuntimeError: Node (ConstantOfShape.11) Op (ConstantOfShape) [ShapeInferenceError] shape input must be 1D tensor

   linspace op中输入Num是0-D tensor的情况需要特殊处理

3. SwinTransformer类模型

   In Node, ("If.0", If, "", -1) : ("Cast.15": tensor(bool),) -> ("p2o.pd_op.set_value_.0.0",) , Error No Op registered for SequenceMap with domain_version of 12 ==> Context: Bad node spec for node. Name: SequenceMap.0 OpType: SequenceMap
   

   opset version遗留问题：存在控制流OP时不能正确获取opset version的问题，导致算子转化失败

4. set_value opset version升级
   因为使用了ONNX的SequenceMap，需要opset version >= 17

5. 获取控制流中opset version时，部分算子的GetMinOpsetVersion中也需要获取TensorInfo信息，因此需要将获取控制流block中算子的输出名称提前到GetCfBlockMinOpsetVersion中

[paddle-bot]

Thanks for your contribution!

[risemeup1]

LG TM