Refactor GPT-J output tracing to use standardized decorators

[burtenshaw]

Applied the overlapping GPT-J refactor from the staged PRs: added _can_record_outputs, moved GPTJModel.forward to decorator-based output capture, switched wrapper forwards to @can_return_tuple, and removed manual hidden-state/attention plumbing. Static validation passed (compileall, make style, git diff --check); pytest could not run in this environment because torch is unavailable and importing transformers also reported missing safetensors package metadata.

Target issue: #43979.

Tests:

• python -m compileall src/transformers/models/gptj/modeling_gptj.py
• python -m pytest tests/models/gptj/test_modeling_gptj.py -q
• PYTHONPATH=src python -m pytest tests/models/gptj/test_modeling_gptj.py -q
• make style
• git diff --check

Source PRs:

• Refactor FNet and CVT output tracing #43996
• [ResNet] Refactor output tracing to decorator-based interface #44007
• Ouptut tracing: Standardizing MobileNetv2 #44013
• Refactor DeBERTa's output tracing interface #44044
• Refactor GPT-J to use standardized output tracing (#43979) #44066
• refactor efficientnet output tracing with @capture_outputs and @can_r… #44072
• Refactor RemBERT to use output tracing decorators #44085
• Refactor SpeechT5 output tracing to standardized output capture #44129
• Refactored vits to match standardized output collection interface #44154
• Refactor gptj output tracing to use standardized decorators #44722

[github-actions]

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

run-slow: gptj

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.