fix(vlm): handle legacy conversation data format and check image in data (#2018) [skip ci]

* fix: handle legacy conversation data format and check image in data

* feat: add test for llama vision

* feat: add max_steps to test

* fix: incorrect indent and return preprocess

* feat: use smaller model and dataset

* chore: add extra config for sharegpt dataset

tests/e2e/test_optimizers.py

@@ -56,6 +56,7 @@ class TestCustomOptimizers(unittest.TestCase):
                 "output_dir": temp_dir,
                 "learning_rate": 0.00001,
                 "optimizer": "optimi_adamw",
+                "max_steps": 5,
                 "lr_scheduler": "cosine",
             }
         )