Specify dataset version used by funeTuning workload (#5736)(#5738)

To workaround this issue when loading dataset when running qloar fineTuning nf4 case:
RuntimeError: Dataset scripts are no longer supported, but found viggo.py

* set use_cache=False in LLama2 fineTuning case

* fix datasets version to v3.6.0 in QLora fineTuning case

Author: 1pikachu

examples/gpu/llm/bitsandbytes/requirements.txt

@@ -3,3 +3,4 @@ transformers==4.51.3
 tf-keras
 accelerate==1.1.1
 peft==0.14.0
+datasets==3.6.0

examples/gpu/llm/fine-tuning/Llama2/llama2_ft.py

@@ -238,7 +238,8 @@ def train():
         model_args.model_name_or_path,
         cache_dir=custom_args.cache_dir,
         attn_implementation="sdpa" if model_args.use_flashattn else "eager",
-        use_safetensors=False
+        use_safetensors=False,
+        use_cache=False
         # num_hidden_layers=1, # set only 1 decoder layer
     )