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.

[atom2-source]

i agree this was abandoned from low usage with no thought put into that devices on edge are far more powerful and capable it was abandoned too early maybe we could do a community fork of a earlier version and serve it ourselves.