* recast loralayer, norm, lmhead + embed token weights per original qlora * try again for the fix * refactor torch dtype picking * linter fixes * missing import for LoraLayer * fix install for tests now that peft is involved
* recast loralayer, norm, lmhead + embed token weights per original qlora * try again for the fix * refactor torch dtype picking * linter fixes * missing import for LoraLayer * fix install for tests now that peft is involved