Only copy external-link index.html for PACKAGE_LINKS_ALLOW_LIST packages (#8115) (#8134)

## Summary

Fixes #8115 — XPU nightly wheels fail to install because runtime deps
such as `tcmlib==1.5.0` disappear from
`https://download.pytorch.org/whl/nightly/xpu/`.

## Root cause

`PACKAGE_LINKS_ALLOW_LIST` packages (the `torch_xpu` deps like `tcmlib`,
plus `numpy`, `nvidia-*`, ...) are link packages whose `index.html`
points at external sources (PyPI CDN / `pypi.nvidia.com`). The
copy-from-parent step copied the parent index into each subdir
**unconditionally**. When the parent index was a relative-path, S3-wheel
index (generated from the wheels that physically live under the subdir),
it overwrote the real links index and dropped externally-hosted versions
such as `tcmlib==1.5.0`.

## Fix

Guard the copy: only copy a links package's `index.html` when it
actually contains external links (`files.pythonhosted.org` or
`pypi.nvidia.com`), and skip it when it only has relative paths.
Everything else is unchanged.

## Test plan

- `python s3_management/manage_v2.py whl/nightly --do-not-upload` and
confirm `whl/nightly/xpu/tcmlib/index.html` retains its external (PyPI)
links incl. `1.5.0`, while `numpy`/`nvidia-*` still propagate into each
`cu*`/`xpu` subdir.

Author: atalman

s3_management/manage_v2.py

@@ -583,6 +583,13 @@ def safe_parse_version(ver_str: str) -> Version:
         return Version("0.0.0")
 
 
+def index_has_external_links(html: str) -> bool:
+    # A links package index must point at external sources (PyPI CDN /
+    # pypi.nvidia.com). A relative-path index is S3-wheel-based and must not
+    # overwrite the real links index. See pytorch/test-infra#8115
+    return "files.pythonhosted.org" in html or "pypi.nvidia.com" in html
+
+
 class S3Index:
     def __init__(self, objects: List[S3Object], prefix: str) -> None:
         self.objects = objects
@@ -1067,6 +1074,19 @@ def upload_package_index(pkg_name: str) -> None:
                         root_obj = BUCKET.Object(key=root_index_key)
                         root_index_html = root_obj.get()["Body"].read().decode("utf-8")
 
+                        # For links packages, only copy an index that points at
+                        # external sources; skip a relative-path (S3 wheel) index
+                        # so it cannot overwrite the real links index. Non-links
+                        # packages are unaffected (they never reach this branch).
+                        if (
+                            pkg_name.lower() in PACKAGE_LINKS_ALLOW_LIST
+                            and not index_has_external_links(root_index_html)
+                        ):
+                            print(
+                                f"INFO: Skipping copy of {root_index_key}: no external links"
+                            )
+                            return
+
                         # Upload to subdirectory in S3
                         BUCKET.Object(key=f"{subdir}/{compat_pkg_name}/index.html").put(
                             ACL="public-read",
@@ -1228,6 +1248,19 @@ def save_pep503_htmls(self) -> None:
                         with open(root_index_path, mode="r", encoding="utf-8") as src:
                             root_index_html = src.read()
 
+                        # For links packages, only copy an index that points at
+                        # external sources; skip a relative-path (S3 wheel) index
+                        # so it cannot overwrite the real links index. Non-links
+                        # packages are unaffected (they never reach this branch).
+                        if (
+                            pkg_name.lower() in PACKAGE_LINKS_ALLOW_LIST
+                            and not index_has_external_links(root_index_html)
+                        ):
+                            print(
+                                f"INFO: Skipping copy of {root_index_path}: no external links"
+                            )
+                            continue
+
                         # Save to subdirectory
                         with open(
                             path.join(pkg_dir, "index.html"),