fix(deps): bump datasets versions for all providers (#3382)

Not doing so results in errors of the kind you see in:
https://github.com/llamastack/llama-stack-ops/actions/runs/17565092546/job/49890264353

Author: ashwinb

llama_stack/providers/registry/datasetio.py

@@ -30,7 +30,7 @@ def available_providers() -> list[ProviderSpec]:
             adapter=AdapterSpec(
                 adapter_type="huggingface",
                 pip_packages=[
-                    "datasets",
+                    "datasets>=4.0.0",
                 ],
                 module="llama_stack.providers.remote.datasetio.huggingface",
                 config_class="llama_stack.providers.remote.datasetio.huggingface.HuggingfaceDatasetIOConfig",
@@ -42,7 +42,7 @@ def available_providers() -> list[ProviderSpec]:
             adapter=AdapterSpec(
                 adapter_type="nvidia",
                 pip_packages=[
-                    "datasets",
+                    "datasets>=4.0.0",
                 ],
                 module="llama_stack.providers.remote.datasetio.nvidia",
                 config_class="llama_stack.providers.remote.datasetio.nvidia.NvidiaDatasetIOConfig",

llama_stack/providers/registry/post_training.py

@@ -48,7 +48,7 @@ def available_providers() -> list[ProviderSpec]:
         InlineProviderSpec(
             api=Api.post_training,
             provider_type="inline::huggingface-gpu",
-            pip_packages=["trl", "transformers", "peft", "datasets", "torch"],
+            pip_packages=["trl", "transformers", "peft", "datasets>=4.0.0", "torch"],
             module="llama_stack.providers.inline.post_training.huggingface",
             config_class="llama_stack.providers.inline.post_training.huggingface.HuggingFacePostTrainingConfig",
             api_dependencies=[