Fix flash attention speed issue (#32028)

Add the lru_cache for speed

src/transformers/utils/import_utils.py

@@ -820,6 +820,7 @@ def is_flash_attn_greater_or_equal_2_10():
     return version.parse(importlib.metadata.version("flash_attn")) >= version.parse("2.1.0")
 
 
+@lru_cache()
 def is_flash_attn_greater_or_equal(library_version: str):
     if not _is_package_available("flash_attn"):
         return False