Keep the PyPI-sourced CUDA deps in PACKAGE_LINKS_ALLOW_LIST only

cuda_python, cuda_bindings, cuda_pathfinder, cuda_toolkit and nvidia_ml_py
appeared in both allow lists. They are sourced from PyPI by
update_dependencies.py, so their index is meant to be copied from the parent
rather than regenerated: that is what #8483 added them to
PACKAGE_LINKS_ALLOW_LIST for. It left them in PACKAGE_ALLOW_LIST, which keeps
manage_v2 generating an index for them from whatever wheels it finds,
competing with the copied one.

They were the only five entries present in both lists; the links list is
untouched, so all five still get copied into the arch subdirectories.

whl/test/cu134 currently advertises cuda-pathfinder and cuda-python in its
listing while both targets 403.

Author: atalman

s3_management/manage_v2.py

@@ -286,12 +286,12 @@
         "flash_attn_3",
         # vllm
         "ninja",
-        "cuda_python",
-        "cuda_bindings",
-        "cuda_pathfinder",
-        "cuda_toolkit",
+        # cuda_python, cuda_bindings, cuda_pathfinder, cuda_toolkit and
+        # nvidia_ml_py are sourced from PyPI by update_dependencies.py and live
+        # in PACKAGE_LINKS_ALLOW_LIST only. Listing them here as well makes
+        # manage_v2 regenerate their index from whatever wheels it finds,
+        # competing with the copied one.
         "pynvml",
-        "nvidia_ml_py",
         "einops",
         "packaging",
         "nvidia_cudnn_frontend",