Fix callback tensor inputs that are never bound in the denoising loop

[LuShadowX]

What does this PR do?

Three pipelines list names in _callback_tensor_inputs that are never bound in their __call__, so callback_kwargs[k] = locals()[k] raises KeyError as soon as a user requests one:

pipeline	name	why it isn't a local
LEditsPPPipelineStableDiffusion	prompt_embeds	never assigned; the matching pop was already commented out
LEditsPPPipelineStableDiffusionXL	negative_add_time_ids	this pipeline has no negative time ids; pop also commented out
StableDiffusionXLControlNetUnionInpaintPipeline	masked_image_latents	mask, _ = self.prepare_mask_latents(...) discards it

On main:

pipe(
    ...,
    callback_on_step_end=lambda p, i, t, kw: kw,
    callback_on_step_end_tensor_inputs=["prompt_embeds"],
)
# KeyError: 'prompt_embeds'

Each allowlist now matches the actual denoising-loop locals. Both LEdits++ test classes are plain unittest.TestCase, so the PipelineTesterMixin callback tests never ran against them — this adds test_callback_inputs to each. StableDiffusionXLControlNetUnionInpaintPipeline has no test file, so it is covered by the fix alone; I found it with an AST sweep over src/diffusers/pipelines for the same defect, which now reports zero remaining instances.

Addresses Issue 4 of #13635 (coordination comment: #13635 (comment)). The commented-out check_inputs() call that issue also mentions is left alone.

Tests

DIFFUSERS_TEST_DEVICE=cpu pytest tests/pipelines/ledits_pp -q
# main:        2 failed, 6 passed, 2 skipped
#              E KeyError: 'prompt_embeds'
#              E KeyError: 'negative_add_time_ids'
# this branch: 8 passed, 2 skipped

DIFFUSERS_TEST_DEVICE=cpu pytest tests/pipelines/controlnet/test_controlnet_inpaint_sdxl.py \
                                tests/pipelines/controlnet/test_controlnet_sdxl.py -q
# 147 passed, 64 skipped

make quality
# All checks passed! / 2020 files already formatted

python utils/check_copies.py && python utils/check_dummies.py && python utils/check_forward_call_docstrings.py
# clean

utils/check_repo.py and utils/check_inits.py fail the same way on a pristine checkout of main on my machine (missing diffusers.models.auto, and a lookup for src/transformers), so they are unrelated to this diff.

Self-review notes

• Fixed the allowlists to match real locals rather than inventing bindings for the missing names — per .ai/AGENTS.md, don't guess intent and silently correct behavior.
• Deleted the two commented-out callback_outputs.pop(...) lines instead of leaving them next to the change, per the no-dead-code rule.
• The new tests reuse the callback_inputs_all check from PipelineTesterMixin.test_callback_inputs rather than inventing a pattern, per .ai/testing.md.
• Left deliberately: in both LEdits++ pipelines negative_prompt_embeds resolves to the __call__ argument, which is None unless the caller passes it — the real uncond tensor is uncond_embeddings in the SD pipeline. It doesn't crash, so I'd rather have your call on it than change it here.
• Left deliberately: SDXL's commented-out check_inputs() call. Re-enabling it changes validation behavior well beyond this bug.

Before submitting

• Did you use an AI agent (Claude Code, Codex, Cursor, etc.) to help with this PR? If so:
  • Did you read the Coding with AI agents guide?
  • Did you run the self-review skill on the diff?
  • Did you share the final self-review notes in the PR description or a comment?
• Did you read the contributor guideline?
• Was this discussed/approved via a GitHub issue? ledits_pp model/pipeline review #13635
• Did you write any new necessary tests?

Who can review?

@yiyixuxu @asomoza

[github-actions]

Hi @LuShadowX, thanks for the PR! It does not appear to link an issue it fixes. If this PR addresses an existing issue, please add a closing keyword (e.g. Fixes #1234) to the PR description so the issue is linked. See the contribution guide for more details. If this PR intentionally does not fix a tracked issue, a maintainer can add the no-issue-needed label to silence this reminder.