[tinker] Fix checkpointing for multi-node runs

[pcmoritz]

Before this fix, there would be an error like

  2026-05-29 20:09:06,716 - ERROR - skyrl: Error saving checkpoint for model model_5b6cb10d, checkpoint <...>: [Errno 2] No such file
  or directory: '/tmp/tmp6kkqkzrg/model'

[gemini-code-assist]

Code Review

This pull request updates checkpoint saving and loading in skyrl_train_backend.py to stage temporary files on the same shared filesystem as the target paths, resolving multi-node deployment issues where remote workers and the engine do not share local /tmp. A review comment identifies a potential FileNotFoundError if reference_path is a relative path, and suggests resolving it to an absolute path using os.path.abspath before extracting the directory name.

[gemini-code-assist]

If reference_path is a relative path (e.g., a simple filename like "checkpoint.tar"), os.path.dirname(reference_path) will return an empty string "". Calling os.makedirs("", exist_ok=True) will then raise a FileNotFoundError: [Errno 2] No such file or directory: ''.

To prevent this and ensure robustness for relative paths, resolve the absolute path using os.path.abspath before extracting the directory name.

Suggested change

	staging_root = str(os.path.dirname(reference_path))
	os.makedirs(staging_root, exist_ok=True)
	return staging_root
	staging_root = os.path.dirname(os.path.abspath(reference_path))
	os.makedirs(staging_root, exist_ok=True)
	return staging_root