Fix UnslothTrainingArguments not patching trl.Config properly

[Erland366]

Few updates ago, we patch trl.Config especially in the bf16 and fp16 args to bypass the Ampere GPU error -> https://github.com/unslothai/unsloth/blob/main/unsloth/models/rl.py#L235-L267

But the patching is not propagate to the inheritance class that is using dataclass. Therefore, we should inherit it using Python Class instead to propagate the patching.

This fixes training error on Mistral CPT