[CustomOp] Support duplicable op input and output

[chenwhql]

PR types

New features

PR changes

OPs

Describe

[CustomOp] Support duplicable op input and output.

This PR add variable length tensor input and output in op, for example:

// add new param type: const std::vector<paddle::Tensor>&
std::vector<paddle::Tensor> ConcatForwardDynamicAxis(
    const std::vector<paddle::Tensor>& inputs, const paddle::Tensor& axis_t) {
  ...
}

// add new param type: std::vector<std::vector<int64_t>>
std::vector<std::vector<int64_t>> ConcatInferShapeDynamicAxis(
    std::vector<std::vector<int64_t>> input_shapes,
    std::vector<int64_t> axis_shape) {
  return {std::vector<int64_t>(input_shapes[0].size(), -1)};
}

// add new param type: std::vector<paddle::DataType
std::vector<paddle::DataType> ConcatInferDtypeDynamicAxis(
    std::vector<paddle::DataType> input_dtypes, paddle::DataType axis_dtype) {
  return {input_dtypes[0]};
}

// the input string of `const std::vector<paddle::Tensor>&` need wrapped by `paddle::Vec()`
PD_BUILD_OP(custom_concat)
    .Inputs({paddle::Vec("X"), "Axis"})
    .Outputs({"Out"})
    .SetKernelFn(PD_KERNEL(ConcatForwardDynamicAxis))
    .SetInferShapeFn(PD_INFER_SHAPE(ConcatInferShapeDynamicAxis))
    .SetInferDtypeFn(PD_INFER_DTYPE(ConcatInferDtypeDynamicAxis));

// note the warp order: paddle::Grad(paddle::Vec("X"))
PD_BUILD_GRAD_OP(custom_concat)
    .Inputs({paddle::Vec("X"), paddle::Grad("Out"), "Axis"})
    .Outputs({paddle::Grad(paddle::Vec("X"))})
    .SetKernelFn(PD_KERNEL(ConcatBackwardDynamicAxis));

[JiabinYang]

LGTM

[JiabinYang]

Bad comments, describe what's for in detail

[chenwhql]

thx, I will remove it in next PR

[JiabinYang]

clang format using for(const auto& x: vec_x) is better?

[chenwhql]

if using for(const auto& x: vec_x), we cannot tell users which tensor cause error in error message.

[Aurelius84]

我们可以统一一下函数的输入参数类型，infershape、inferDatatype可以都使用const &，更符合编程习惯和直觉。可以在下个PR修改

[chenwhql]

thx，下个PR修改