✨[Feature] Support ShapeTensor I/O

[peri044]

Is your feature request related to a problem? Please describe. Currently if a network takes an input tensor which is used as shape for other layers in the graph, a shape optimization profile has to be set for this input. Eg: In the case of dynamic shapes + fallback, sym_int layers fallback to Pytorch which are passed as inputs to other subgraphs. The min, opt, max values for sym_int inputs need to be set at compile time for TensorRT to pass.

Describe the solution you'd like

Describe alternatives you've considered

Additional context