All of our fine-tuning tests were getting skipped because our check for cuda was too strict, we only wanted to do that if a quantization_config was specified. This was missed because the tests were too complicated and had a lot of logic in them.

This PR refactors our existing LLM fine-tuning tests by splitting them into two:

1. Without Quantization
2. With Quantization

All of the without quantization tests will run going forward, and the quantization tests will be skipped if no GPUs are available.

What's included in the non-quantization tests (FYI, none of these were running until now, they were all getting incorrectly skipped):

1. Full Fine-tuning
2. LoRA fine-tuning using defaults
3. LoRA fine-tuning using modified parameters (NEW)
4. AdaLora fine-tuning using defaults (NEW)
5. AdaLoRA fine-tuning using modified parameters (NEW)
6. Adaption Prompt using defaults
7. Adaption prompt using modified parameters (NEW)

It also sets the pytest mark once at the top of the file since we were not being consistent with adding this marker throughout.