Fix various bugs for LoRA training

[AstrisCantCode]

• Allow mistral models to have all modules targeted selectively, since the module names are the same as those for llama models
• Fixed issue that occurred because prepare_model_for_kbit_training was being run on unquantized models that didn't need it (causing an OOM error)
• Fixed issue in the training arguments that didn't allow BF16 training to take place even when the appropriate options were selected in the model menu
• Fixed typo that erroneously (always) enabled IPEX

Checklist:

• I have read the Contributing guidelines.

[oobabooga]

Thank you for the fixes