Fix PPFormulaNet training-loss double-shift in encoder-decoder forward

[ShamSaleem]

What does this PR do?

Fixes #47317

PPFormulaNetForConditionalGeneration.forward() computes its loss with
self.loss_function() → ForCausalLMLoss, which shifts labels left by one
(labels[..., 1:]) when shift_labels is not supplied. PPFormulaNet is an
encoder-decoder model, so its logits are already aligned with labels and
no shift is needed. The extra shift drops the last target token and trains the
model against misaligned labels.

This passes shift_labels=labels to ForCausalLMLoss, which takes the
"already-shifted" path and leaves the labels aligned with the logits.

This is the same fix already merged for Florence2 in #46898 (and Moonshine
before it). PPFormulaNet's modular model inherits from
Florence2ForConditionalGeneration but overrides forward(), so the
Florence2 fix never propagated to it — eliaghazal flagged PPFormulaNet as
likely-affected in the #47090 discussion.

Change is applied to modular_pp_formulanet.py (source of truth) and the
regenerated modeling_pp_formulanet.py.

Before submitting

• This PR fixes a typo or improves the docs.
• Did you read the contributor guideline and the PR checks?
• Was this discussed/approved via a Github issue? Yes — PPFormulaNet training-loss bug #47317.
• Did you make sure to update the documentation?
• Did you write any new necessary tests? — No new test, matching the merged precedent Fix Florence2 training-loss double-shift (same pattern as Moonshine #… #46898 (source-only); happy to add a loss-alignment regression test if preferred.

Note: AI assistance was used to help locate and draft this fix.

Who can review?

@zucchini-nlp (reviewed & merged the identical Florence2 fix #46898)

[zucchini-nlp]

Already being fixed on another PR, closing as duplicate