Fix path-traversal vulnerability in emergency P2P checkpoint service

[YuvalElbar6]

A malicious or compromised peer on the P2P network could supply a manifest whose rel_path contained '..' segments or an absolute path, causing P2PNode.fetch_shard_from_peer() to write attacker-controlled bytes outside the staging directory (e.g. a .pth file in site-packages, yielding persistent RCE on the training host).

• Add _safe_path_join() which joins a peer-supplied relative path onto a base directory only if the resolved result stays inside that base. Resolution goes through os.path.realpath so symlink-escape attempts are caught as well.
• Apply the helper on both sides of the wire:
  • Client: fetch_shard_from_peer() validates every manifest entry against stage_dir and aborts the whole fetch on any unsafe entry.
  • Server: handle_download() replaces the substring '..' check with the same resolve-based containment check against self.directory.
• Log every rejection with peer and request context.
• Add regression tests for the helper and both call sites.

Reported via the Google OSS VRP.