[VLMs] fix flash-attention tests (#37603)

* fix one test

* fa2 ln test

* remove keys from config recursively

* fix

* fixup

tests/models/llava_next/test_modeling_llava_next.py

@@ -331,10 +331,6 @@ class LlavaNextForConditionalGenerationModelTest(ModelTesterMixin, GenerationTes
     def test_training_gradient_checkpointing_use_reentrant_false(self):
         pass
 
-    @unittest.skip("FlashAttention only support fp16 and bf16 data type")
-    def test_flash_attn_2_fp32_ln(self):
-        pass
-
     @unittest.skip(
         "VLMs need lots of steps to prepare images/mask correctly to get pad-free inputs. Can be tested as part of LLM test"
     )