feat(trainer): log individual losses from loss_dict

[Abdeltoto]

What does this PR do?

Fixes #31081.

When a model returns auxiliary losses alongside the main loss (e.g. via a
loss_dict field in its ModelOutput), the Trainer currently only logs the
combined loss. That makes debugging multi-term objectives painful: you can
see total loss going down without knowing which term is actually moving.

This PR teaches the Trainer to also log each scalar term it finds in
outputs.loss_dict, plus any top-level *_loss scalar attribute on the
output, under namespaced keys like loss_dict_<name> and loss_<name>.
Behaviour is opt-in by virtue of the model itself: if no extra losses are
returned, nothing changes. The main loss value and all existing logs are
unchanged.

Implementation notes

• New buffer self._aux_losses_accumulator on the Trainer, mirroring how
  _total_loss_scalar already works.
• In training_step, after compute_loss(..., return_outputs=True), scalar
  tensor entries from outputs.loss_dict and from outputs.<name>_loss are
  detached, gradient-accumulation-scaled, and accumulated. Same DP mean and
  same num_items_in_batch normalization as the main loss, so the numbers
  are comparable.
• In _maybe_log_save_evaluate, accumulators are gathered across processes
  with nested_gather (consistent with the main tr_loss path), averaged
  over the logging window, and added to logs. Then the buffers are reset.
• Non-tensor / non-scalar values are silently ignored, so models that put
  arbitrary metadata in loss_dict won't crash the Trainer.

No public API change. No new dependency. The diff is mostly localized to two
methods in trainer.py.

Tests

tests/trainer/test_trainer.py::TrainerIntegrationTest::test_trainer_logs_auxiliary_losses_from_loss_dict

A small RegressionPreTrainedModelWithLossDict (added to
tests/trainer/trainer_test_utils.py) returns loss, loss_dict={'mse', 'l1'}, and a top-level extra_loss. The test runs a tiny Trainer.train()
and asserts the resulting log entries contain the expected
loss_dict_mse, loss_dict_l1, and loss_extra keys with finite,
positive values. Ran locally:

Ran 1 test in 0.245s
OK

Code Agent Policy

• I confirm that this is not a pure code agent PR.

I used Cursor as a coding assistant (the commit trailer says
Made-with: Cursor), but I read every diff, ran the tests locally, wrote
the description myself, and own the change. Happy to iterate on review
feedback.

Before submitting

• This PR fixes a typo or improves the docs.
• Did you read the contributor guideline?
• Was this discussed/approved via a GitHub issue? See Log multiple losses used along with the combined losses when a model returns a dictionary of losses. #31081.
• Did you make sure to update the documentation with your changes? No
  user-facing docs change — the new keys appear automatically when a
  model already returns loss_dict.
• Did you write any new necessary tests?

Who can review?

@SunMarc — Trainer maintainer per the template.

[Rocketknight1]

Sorry, we really don't want code agent PRs on old issues like this!