[mtthw13]

Replaces manual output_attentions/output_hidden_states/return_dict boilerplate in GPT-Neo with the hook-based decorator system.

Changes:

• Added _can_record_outputs = {"hidden_states": GPTNeoBlock, "attentions": GPTNeoAttention} on GPTNeoPreTrainedModel
• Added @capture_outputs + @merge_with_config_defaults on GPTNeoModel.forward
• Added @can_return_tuple on GPTNeoForCausalLM, GPTNeoForSequenceClassification, GPTNeoForTokenClassification, GPTNeoForQuestionAnswering
• Dropped output_attentions, output_hidden_states, return_dict from all forward signatures
• Removed parameter resolution lines and manual collection loops
• GPTNeoBlock now returns hidden_states directly (not a tuple)
• Attention always returns (attn_output, attn_weights)
• Removed unused imports: BaseModelOutputWithPastAndCrossAttentions, CausalLMOutputWithCrossAttentions
• Updated test: removed output_attentions=True from direct GPTNeoSelfAttention call in test_local_attn_probs

Tests: All 111 GPT-Neo model tests pass.

[mtthw13]

Fixes #43979.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: gpt_neo

[ArthurZucker]

Thanks for picking this up from #43979! That call to contributions has been closed, and the reference refactor (#43590) landed afterwards with a slightly different pattern, so this branch is out of date and I'm closing it. If you'd still like to migrate GPT-Neo, a fresh PR following #43590 is welcome.