[compiler] Support Shape Inference for determined by input data

[icodo98]

What

Let's support Shape Inference for determined by input data

Why

There are some operations(ex. Reshape, Pad) which determine output shape by input tensor data.
Current ShapeInference only allows access to input data when input is CircleConst.
It leaves output shape unknown or stop inferencing when input is CircleNode instead CircleConst.
However, there are some operations (ex. PACK) that determines output data in compile time.
Maybe we can inference output shape for these cases.

[icodo98]

It leaves output shape unknown or stop inferencing when input is CircleNode instead CircleConst.

• Reshape case
  • [compiler] Support dynamic shape for reshape  #13927 (comment)

  When reshape/shape is CircleNode

  • use shape_by_input & output_shape would be dynamic
  • Example case : reshape(input, shape_node) --> [?, ?, ?]

With principal above, .opt.circle is different from .tflite in YamNet

After ShapeInferencePass output shape of Reshape will be [?, ?, ?, ?].