Updated automatic model training notebook for "high performance" models

[Jyers]

I recently started working on training some higher performance models using the advanced model training notebook.
During this process, I ended up rewriting some chunks of the notebook to implement the recommendations provided for training higher quality models. As well as squashing a few existing bugs while I was at it.
I figured I'd share it here for anyone who's looking for a pre-made solution for training better models.
@dscripka if you think it would be beneficial, I can open a PR to add it alongside the other notebooks. Otherwise, I'm sure people will be able to find it here if they're looking.

Changes:

• Unified procedures for downloading full IR, Audioset, and FMA datasets
• Avoid soundfile FLAC bitflip issue #163 and fsspec.exceptions.FSTimeoutError when downloading dataset huggingface/datasets#7164 during downloads
• Implemented new onnx to tflite from Update onnx to tflite conversion to remove onnx-tf #75

I was also thinking about making a quick writeup on how to easily run the notebook using WSL. As that might help any windows users who want to train models for longer than the free Google Colab runtimes allow. If anyone is interested in that let me know!

model_training_notebook.zip