[Model] Add user-configurable task for models that support both generation and embedding (#9424)

tests/models/decoder_only/vision_language/test_phi3v.py

@@ -89,6 +89,7 @@ def run_test(
 
     # max_model_len should be greater than image_feature_size
     with vllm_runner(model,
+                     task="generate",
                      max_model_len=4096,
                      max_num_seqs=2,
                      dtype=dtype,