Propagate onnx.dim_params in shape inference

[chentong319]

ONNX model provides symbolic name for dynamic dimension of input tensor with argument attribute "onnx.dim_params". This PR tries to propagate this symbolic name for dynamic dimension during shape inference with extra attribute to operation "onnx.dim_params_[n]", where n is the index of the output tensor.
The major change is on IndexExpr to carry such info, if it exists. Other changes include the use of the dim_param info in shape inference and test case.
So far, the usu of dim_param info is far from complete. The PR is kind of experimental.
At least one benefit of this approach is that the readability of the output of onnx-mlir can be improved with the symbolic names.
Another potential benefit is the reuse the existing shape inference code to do symbolic dim propagation. For example, without touching the shape inference for onnx.MatMul, we can get the result of the following test case:

func.func @test_matmul_2_param(%arg0 : tensor<16x?x64x42xf32> {onnx.dim_params="1:dim1"}, %arg1 : tensor<42x?xf32> {onnx.dim_params="1:dim2"}) -> tensor<*xf32> {
  %0 = "onnx.MatMul"(%arg0, %arg1) : (tensor<16x?x64x42xf32>, tensor<42x?xf32>) -> tensor<*xf32>
  "onnx.Return"(%0) : (tensor<*xf32>) -> ()
}

Result from onnx-mlir-opt --shape-inference

  func.func @test_matmul_2_param(%arg0: tensor<16x?x64x42xf32> {onnx.dim_params = "1:dim1"}, %arg1: tensor<42x?xf32> {onnx.dim_params = "1:dim2"}) -> tensor<16x?x64x?xf32> {
    %0 = "onnx.MatMul"(%arg0, %arg1) {onnx.dim_params_0 = "1:dim1,3:dim2"} : (tensor<16x?x64x42xf32>, tensor<42x?xf32>) -> tensor<16x?x64x?xf32>
    onnx.Return %0 : tensor<16x?x64x?xf32>
  }

[tungld]

Suggestion: how about using onnx.out_dim_params_? I thought onnx.dim_params_0 was for the first input, but actually it was for the first output.

[tungld]

LGTM with a minor change in naming.

The propagation looks effective. Let's see how we can utilize this information in future PRs.

For the changes to IndexExpr, I would like to hear comments from @AlexandreEichenberger as well.