fix: Post Training Model change in Tests in order to make it less intensive (#2991)

# What does this PR do?

Changed from` ibm-granite/granite-3.3-2b-instruct` to`
HuggingFaceTB/SmolLM2-135M-Instruct` so it as not resource intensive in
CI

Idea came from -
https://github.com/meta-llama/llama-stack/pull/2984#issuecomment-3140400830

tests/integration/post_training/test_post_training.py

@@ -70,7 +70,6 @@ class TestPostTraining:
         ],
     )
     @pytest.mark.timeout(360)  # 6 minutes timeout
-    @skip_because_resource_intensive
     def test_supervised_fine_tune(self, llama_stack_client, purpose, source):
         logger.info("Starting supervised fine-tuning test")
 
@@ -111,7 +110,7 @@ class TestPostTraining:
         # train with HF trl SFTTrainer as the default
         _ = llama_stack_client.post_training.supervised_fine_tune(
             job_uuid=job_uuid,
-            model="ibm-granite/granite-3.3-2b-instruct",
+            model="HuggingFaceTB/SmolLM2-135M-Instruct",  # smaller model that supports the current sft recipe
             algorithm_config=algorithm_config,
             training_config=training_config,
             hyperparam_search_config={},