Feat/kaggle-gguf-on-tmp

[Erland366]

Now we move all GGUF Process to /tmp so user can convert GGUF using Kaggle

We manage to save the GGUF on /tmp but now the problem is how user will download the GGUF since you can't access /tmp through GUI and FileLink seems not working .-.

Also fixing #1222 since HFFileSystem is hardcoded to use linux format of forward slash instead of adapting to the user's OS

Here's my run for GGUF on Mistral 7b (I already fix the wrong name on the print, I am to lazy to rerun it again LOL)

Because the save_pretrained_gguf still not "solved" (need to discuss how to approach the download process). I put it as draft first

[danielhanchen]

@Erland366 If people cannot download it, thats OK - the goal is to make uploading 16bit merged weights possible - GGUF is also the same - the conversion should default to the tmp directly and save it into that directly