[onnx->torch] The folding behavior of the reducemax operator is inconsistent for float and int.

[chenxinyi0415]

For two test

func.func @test_reduce_max_empty_set_fp(%arg0: !torch.vtensor<[2,0,4],f32>, %arg1: !torch.vtensor<[1],si64>) -> !torch.vtensor<[2,1,4],f32> attributes {torch.onnx_meta.ir_version = 9 : si64, torch.onnx_meta.opset_version = 20 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
  %0 = torch.operator "onnx.ReduceMax"(%arg0, %arg1) {torch.onnx.keepdims = 1 : si64} : (!torch.vtensor<[2,0,4],f32>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[2,1,4],f32>
  return %0 : !torch.vtensor<[2,1,4],f32>
}

func.func @test_reduce_max_empty_set_int(%arg0: !torch.vtensor<[2,0,4],si32>, %arg1: !torch.vtensor<[1],si64>) -> !torch.vtensor<[2,1,4],si32> attributes {torch.onnx_meta.ir_version = 9 : si64, torch.onnx_meta.opset_version = 20 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
  %0 = torch.operator "onnx.ReduceMax"(%arg0, %arg1) {torch.onnx.keepdims = 1 : si64} : (!torch.vtensor<[2,0,4],si32>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[2,1,4],si32>
  return %0 : !torch.vtensor<[2,1,4],si32>
}

run

../torch-mlir/bin/torch-mlir-opt --torch-onnx-to-torch-backend-pipeline test.mlir

Will both conversion to fullOp from onnx to torch.
For int type , fullOp will be folded to torch.vtensor.literal. But for float type, will not be folded because of this code

if (isa<IntegerType>(elementType)) {
    int64_t value = 0;
    if (matchPattern(getFillValue(), m_TorchConstantInt(&value))) {
      Attribute attribute = IntegerAttr::get(elementType, value);
      return DenseElementsAttr::get(shapedty, attribute);
    }
  }
  if (isa<FloatType>(elementType)) {
    double value = 0.0;
    if (matchPattern(getFillValue(), m_TorchConstantFloat(&value))) {
      Attribute attribute = FloatAttr::get(elementType, value);
      return DenseElementsAttr::get(shapedty, attribute);
    }
  }

If change FloatType to mlir::FloatType will successfully be folded, maybe these is a bug, we need to fix it? Because the namespace of mlir and torch both have FloatType.