cache_utils: fix QuantizedLayer to correctly propagate reorder_cache, crop, and batch ops to quantized buffers

[GitGlimpse895]

What does this PR do?

QuantizedLayer maintains two separate storage regions: a full-precision
residual buffer (self.keys / self.values) and a quantized buffer
(self._quantized_keys / self._quantized_values). However, the four
mutation methods inherited from DynamicLayer — reorder_cache,
crop, batch_repeat_interleave, and batch_select_indices — only
operated on the residual buffer, silently leaving the quantized buffer
untouched.

Concrete failure modes:

• Beam search (reorder_cache): the quantized buffer stays in
  original beam order while the residual reorders, causing crossed
  attention across beams with no error raised.
• Constrained generation rollback (crop): cumulative_length
  diverges from the actual stored state, corrupting subsequent
  get_seq_length calls.
• Group beam search / contrastive decoding (batch_select_indices,
  batch_repeat_interleave): batch dimension of the quantized buffer
  is never updated, producing mismatched batch sizes between the two
  storage regions.

This PR overrides all four methods in QuantizedLayer. Since
_quantized_keys/_quantized_values are opaque backend objects for
both QuantoQuantizedLayer and HQQQuantizedLayer, the fix uses a
dequantize → operate → re-quantize pattern, which is backend-agnostic
and does not compound quantization error meaningfully beyond what is
already introduced at the original quantization step.

Fixes # (issue)

Before submitting

• This PR fixes a typo or improves the docs (you can dismiss the other checks if that's the case).
• Did you read the contributor guideline,
  Pull Request section?
• Was this discussed/approved via a Github issue or the forum? Please add a link
  to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?

Who can review?

@gante @SunMarc

• I confirm that this is not a pure code agent PR.

[zucchini-nlp]

tbh haven't seen any requets to add beam search support with quantized cache prev, do you have a need for that or is this an improvement PR?

[zucchini-nlp]

If you need this feature, I don't mind adding it. But we need a test in tests/generation/test_utils.py. Otherwise, I'd prefer to not increase maintenance cost for things that are "nice to have but not used"

[GitGlimpse895]

Hi @zucchini-nlp, thanks for the feedback!

You’re right—this was primarily an improvement PR and isn't a strict requirement for my current workflow. Since it falls into the "nice to have" category and I definitely understand the goal of keeping maintenance costs down, I’ll go ahead and close this one out.

Thanks again for your time and the review!

[github-actions]

View the CircleCI Test Summary for this PR:

https://huggingface.co/spaces/transformers-community/circle-ci-viz?pr=45510&sha=3a04d2