[Tinker API] Restore variable-length result order

[yapdianang]

Summary

• Return each variable-length training result in the same order as its input example.
• Remove results produced only for padding.

Before

Token-based batching returned results in packed microbatch order. A client could pair one example's mask with another example's log probabilities and fail before the optimizer step.

After

SkyRL maps each result back to the original example index before it returns loss_fn_outputs. The change applies to Trajectory custom GSPO and any other client that reads per-example results.

Testing

uv run --isolated --extra dev --extra tinker --extra skyrl-train \
  pytest tests/backends/skyrl_train/test_token_based_batching_utils.py -q

18 passed.

[gemini-code-assist]

Code Review

This pull request introduces support for the GSPO loss function, including batch normalization, row selection, and pruning metrics. It also implements sequence-based scheduling using sequence IDs to ensure requests are processed in order, alongside corresponding API validation and unit tests. The review feedback highlights three key issues: a potential ValueError when zipping outputs if per_sample_outputs is empty, a potential permanent hang during sequence bootstrapping if the database is reset and a non-zero sequence ID is used, and a potential AttributeError if metrics is None in pipeline parallel setups.

[erictang000]

some initial comments, need to look closer at the seq_id/engine logic changes as well!

[cursor]

Cursor Bugbot has reviewed your changes using default effort and found 1 potential issue.

^{Reviewed by Cursor Bugbot for commit 0e3f075. Configure here.}