[V1] Support any head size for FlexAttention backend (#20467)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

tests/models/test_initialization.py

@@ -22,7 +22,8 @@ def test_can_initialize(model_arch: str, monkeypatch: pytest.MonkeyPatch):
     model_info.check_transformers_version(on_fail="skip")
 
     # FIXME: Possible memory leak in the previous tests?
-    if model_arch == "GraniteSpeechForConditionalGeneration":
+    if model_arch in ("GraniteSpeechForConditionalGeneration",
+                      "KimiVLForConditionalGeneration"):
         pytest.skip("Avoid OOM")
 
     # Avoid OOM and reduce initialization time by only using 1 layer