Upgrade transformers to 5.5.3 and refactor hf_transformers_utils into subpackage (#21569)

python/pyproject.toml

@@ -35,6 +35,7 @@ dependencies = [
   "modelscope",
   "msgspec",
   "ninja",
+  "easydict",  # Required by remote model code (e.g. DeepSeek-OCR) loaded via trust_remote_code; validated by transformers 5.4+ check_imports
   "numpy",
   "nvidia-cutlass-dsl>=4.4.1",
   "nvidia-ml-py",
@@ -70,8 +71,8 @@ dependencies = [
   "av ; sys_platform == 'linux' and (platform_machine == 'aarch64' or platform_machine == 'arm64' or platform_machine == 'armv7l')",
   "torchvision",
   "tqdm",
-  "mistral_common>=1.9.0",
-  "transformers==5.3.0",
+  "mistral_common>=1.11.0",
+  "transformers==5.5.4",
   "uvicorn",
   "uvloop",
   "watchfiles",