[SFT] Treat past_key_values and attentions as optional on the model output

[Hakureirm]

What does this PR do?

_chunked_ce_forward reads outputs.past_key_values and outputs.attentions
unconditionally when it rebuilds the output, so a model whose forward returns
neither raises AttributeError on the first step. loss_type defaults to
"chunked_nll", so SFTTrainer fails out of the box for those models.

Neither field is ever populated here — the backbone is called with
use_cache=False and attentions are never requested — so they are passed as
None. Measured on trl-internal-testing/tiny-Qwen2ForCausalLM-2.5, at the
point the output is rebuilt, both are already None today. last_hidden_state
and hidden_states, the fields the chunked path actually reads, are present on
every output type involved.

Ten entries of MODEL_FOR_CAUSAL_LM_MAPPING_NAMES (transformers 5.14.1) declare
a return type without past_key_values — mamba, mamba2, falcon_mamba,
xlstm and rwkv keep the recurrence under cache_params or state, xlnet
under mems, and openai-gpt, recurrent_gemma, reformer and xlm have no
state field at all.

Test. trl-internal-testing/tiny-FalconMambaForCausalLM is added to
TestSFTTrainer::test_train, which uses a default SFTConfig and therefore the
default loss_type; it fails on main and passes here, on CPU and on an A800.

No behaviour change for models that carry the fields. Training
trl-internal-testing/tiny-Qwen2ForCausalLM-2.5 for 3 epochs under a fixed seed
on CPU gives byte-identical per-step losses and parameter sums with and without
this change.

tests/test_sft_trainer.py on this branch: 112 passed / 153 skipped on CPU,
187 passed / 78 skipped on an A800, no failures or errors either way.

Fixes #6685

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline, Pull Request section?
• Was this discussed/approved via a GitHub issue? Please add a link to it if that's the case. — SFTTrainer's default chunked_nll path fails on models whose output has no past_key_values #6685
• Did you make sure to update the documentation with your changes? — checked; this is an internal fix with no user-facing API or behaviour change to document.
• Did you write any new necessary tests?

AI writing disclosure

• No AI usage: the PR was written entirely by a human.
• AI-assisted: some parts were suggested or improved by AI, but the PR was written and reviewed by a human.
• AI-generated: the PR was mostly or fully generated by an AI tool.

[cursor]

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

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

Want higher recall? High effort reviews run extra passes and find more bugs. A team admin can switch effort levels in the Cursor dashboard.

^{Reviewed by Cursor Bugbot for commit eb54dda. Configure here.}