TDA4VM: TI qat train script

When doing QAT with xnn.quantize in edgeai-torchvision, weights are discretized (internally) and compensated for the accuracy loss due to that discretization. Although, the weights saved in the checkpoint are floating point, it has been compensated for the error that will occur when quantized.

When TIDL reads the QAT model, it would apply the same kind of discretization for quantization. Please try the QAT model in TIDL or OSRT (for example ONNXRuntime with TIDL offload) - it should give good accuracy.

For a similar discussion, please see this thread:

https://github.com/TexasInstruments/edgeai-torchvision/issues/7