Add TensorRT optimization profiles - #36660

This PR adds optimization profile handling to TF-TRT and enables inference using the experimental dynamic shapes mode (where dynamic shapes mode = explicit batch mode with dynamic input shapes).

This PR is not supposed to change the existing behavior in implicit batch mode.

For each unknown input dimension, a shape optimization profile describes what is the expected [min, max] range for that dimension, and also define an optimum value. "The builder selects the kernels that result in the lowest runtime for the optimum input tensor dimensions, and are valid for all input tensor sizes in the valid range between minimum and maximum dimensions."

To create an engine with dynamic shapes one must define at least one optimization profile. In the current PR we define only a single optimization profile, using the shape of the concrete input data during the first inference call.

The real advantage of dynamic shapes is to have a single engine that can execute the network with different input shapes. Unfortunately it is not possible to demonstrate this with the current PR. Although the TrtShapeOptimizationProfile class is designed to manage multiple profiles, we still need to enable profile generation mode during build phase, to take the full advatage of dynamic shapes mode. That feature is implemented in PR #36080.