fix: retry a bad data download, keep the site step

The buildbot answered once with something that was not ScummVM.zip. refresh_data_dirs gave up at once, Generate the site was skipped, and 228 site tests failed in its place. A failed or unreadable download is now tried three times, with waits of 5 and 20 s, and an empty archive still fails at once. The site step runs after a failed fetch, which stays red on its own step.

.github/workflows/validate.yml

@@ -276,8 +276,11 @@ jobs:
 
       # The pages are gitignored, and the tests that read them fail
       # in CI without it. The suite runs whatever this step returned, so a
-      # generator that stops still leaves every other test reporting.
+      # generator that stops still leaves every other test reporting. It runs
+      # after a failed data fetch too: one bad mirror stays red on its own
+      # step instead of skipping the site and failing 228 site tests with it.
       - name: Generate the site
+        if: ${{ !cancelled() }}
         run: python scripts/generate_site.py
 
       - name: Run test suite