onnxruntime-training web llama example

[bil-ash]

@carzh I know that onnxruntime-training has been deprecated since version 1.19.2 , however, the on-browser training code looks super useful. I looked into the code for training artifact generation for the browser MNIST demo and the desktop C# BERT demo and tried to adapt it for llama(https://huggingface.co/HuggingFaceTB/SmolLM2-135M-Instruct) but was unsuccessful in doing so. So, I am requesting you to provide a detailed example.
It would be kind of you, if you could put up a separate repo showcasing all the steps of in-browser training-> offline artifacts generation, in-browser training, merging trained params into the in-browser model as well as sending the trained params to server and merging it on the server. I guess it will be a nice pre-Christmas project.