[kocean9-freedom]

What does this PR do?

Preserve a one-element bias vector when reversing FuseBiasInLinear. module.weight[:, -1] already has shape [out_features]; the extra squeeze() turns it into a scalar when out_features=1. The restored layer's state dict then fails to load into the original nn.Linear, with size mismatch for bias: torch.Size([]) ... torch.Size([1]). Single-output classification/regression heads, including text rerankers, are affected.

Remove only that redundant squeeze. Multi-output and biasless layers keep their behavior; no dependency or API changes.

Validation on Python 3.12 / PyTorch 2.10 / Transformers 5.19:

• New download-free native FX tests: 6 failures and 8 passing controls on upstream; 14 passed after the fix. They cover one/multiple outputs, 1D/2D/3D inputs, float32/float64, biasless controls, two fuse/reverse cycles, exact parameter shapes/values, strict checkpoint loading, and input/parameter gradients.
• The actual single-output head from cached pretrained cross-encoder-testing/reranker-bert-tiny-gooaq-bce-STv54 reproduces the checkpoint-loading failure on upstream. With the fix, CPU and MPS paired-text scores, strict state loading, model save/reload, gradients, and one SGD update match the untransformed reference. These checks also pass from an isolated built wheel.
• Full-repository Black and Ruff checks, git diff --check, sdist and wheel builds pass. The complete downloaded-model FX/Optimum test suite and CUDA were not run.

The shared-call fix in #2538 is a separate issue and its current head retains this squeeze. In a separate checkout, combining its patch with this one-line fix passes all 24 shared-call and single-output cases. That patch is not included here. The dtype issue in #2533 is also outside this change.

Before submitting

• Added regression tests and checked the real model use case.
• Public arguments and documented behavior are unchanged; no documentation update is needed.
• Ran code style checks and recorded verification limits above.

AI assistance: implementation, validation, review, and PR preparation used OpenAI Codex.

Additional legacy source compatibility

A separate checkout combining #2535, #2539 and #2540 passes 33 offline cases with Transformers 4.57.6 on both PyTorch 2.10.0 and 2.5.1: the two new 14-case suites, #2535's three layouts, and two existing offline helper tests. This is a source compatibility overlay using Hub 0.36.2, which does not satisfy current main's Hub >=1.31 requirement; repository dependency pins are unchanged. It is not a fully resolved legacy installation or a downloaded-model/full-suite claim.