Properly compute files hashes during models publication

This reverts commit ceb2adbc.

The first solution to this problem was to simplify the models archive creation. The hash computation was failing because folders were added to the file_list. We removed these from the list and it seemed good enough. However, we lost the relative path computation that is absolutely necessary when the workers will download the files. In the Arkindex CLI, it's enough because the path we give are already relative ones.

I went back and kept the relative paths computation and only added some filtering.

arkindex_worker/worker/training.py

@@ -55,7 +55,9 @@ def create_archive(path: DirPath) -> Tuple[Path, Hash, FileSize, Hash]:
         for p in path.glob("**/*"):
             x = p.relative_to(path)
             tar.add(p, arcname=x, recursive=False)
-            file_list.append(p)
+            # Only keep files when computing the hash
+            if x.is_file():
+                file_list.append(x)
 
     # Sort by path
     file_list.sort()