Fix(SAM2): mask post-processing with Transformers 5.x

[hlista]

Summary

Fixes SAM2 segmentation when using the current Transformers SAM2 processor.

Sam2Processor no longer returns reshaped_input_sizes in its processed inputs, but the InvokeAI SAM2 pipeline attempts to access that field during mask post-processing

Removing the obsolete argument allows SAM2 mask post-processing to complete successfully.

Processing an image with the SAM2 processor:

processed_inputs = processor(images=image, return_tensors="pt")
print(processed_inputs.keys())

returns:

pixel_values
original_sizes

There is no reshaped_input_sizes

Related Issues / Discussions

Closes #9551

QA Instructions

Tested locally with:

• InvokeAI 6.14.0
• Transformers 5.5.4
• facebook/sam2.1-hiera-large

Before the change, SAM2 segmentation consistently failed with
KeyError: 'reshaped_input_sizes'

After the change, SAM2 segmentation completes successfully

Merge Plan

Checklist

• The PR has a short but descriptive title, suitable for a changelog
• Tests added / updated (if applicable)
• ❗Changes to a redux slice have a corresponding migration
• Documentation added / updated (if applicable)
• Updated What's New copy (if doing a release after this PR)

[Pfannkuchensack]

Findings

No correctness defect found in the change itself. The one-line removal is verified correct against the pinned dependency. Two coverage/accuracy issues remain.

• Medium (test gap): invokeai/backend/image_util/segment_anything/segment_anything_2_pipeline.py:101-104 — this PR fixes a third-party contract drift (Sam2Processor dropped reshaped_input_sizes from its output) and adds no test. There is no test anywhere covering either SAM pipeline: tests/backend/image_util/ contains only test_color_conversion.py and test_vendor_mutable_defaults.py, and a repo-wide search for *segment* / *sam* test files returns nothing. The identical construct still lives in the SAM1 path at invokeai/backend/image_util/segment_anything/segment_anything_pipeline.py:92, so the next upstream transformers bump that drops reshaped_input_sizes from SamProcessor reproduces issue [bug]: Segment Anything V2 fails every time #9551 verbatim with zero CI signal. The failure mode is a runtime AttributeError inside an invocation, i.e. a user-visible generation failure, not an import-time or type error, so neither lint nor typecheck catches it.
  To expose this issue, add a test that runs SegmentAnything2Pipeline.segment() against a randomly initialized Sam2Model(Sam2Config()) plus a real Sam2Processor(image_processor=Sam2ImageProcessor()) and asserts the returned tensor is torch.bool with shape [num_masks, 3, original_height, original_width]. This needs no downloaded weights and runs in seconds. The same test shape should be added for SegmentAnythingPipeline so the still-present reshaped_input_sizes access in the SAM1 path is pinned by a test rather than by luck.

• Low (stale, now actively misleading comment): invokeai/backend/image_util/segment_anything/segment_anything_2_pipeline.py:6 — the comment reads # Import SAM2 components - these should be available in transformers 4.56.0+. After this change the code depends on the 5.x processor output contract (no reshaped_input_sizes), and pyproject.toml:67 pins transformers>=5.5,<5.6. The comment advertises support for a version range the patched code no longer targets. Relatedly, the constructor docstring at invokeai/backend/image_util/segment_anything/segment_anything_2_pipeline.py:22 claims the processor "can be Sam2Processor or Sam2VideoProcessor", but the type annotation is Sam2Processor and invokeai/app/invocations/segment_anything.py:127 only ever constructs Sam2Processor. Both should be corrected while the file is open; this is a documentation-accuracy issue, not a runtime defect.

• Ruff formats is needed.