Add image-text contrastive loss to Blip2ForImageTextRetrieval

[xquantize]

What does this PR do?

Blip2ForImageTextRetrieval already computes ITC and ITM logits, but stops there — there's no loss path, so the class can't be used for training or fine-tuning. This adds the contrastive (ITC) half of that, as discussed in #34019.

Two things are needed for a usable ITC loss:

1. A learnable temperature. The original LAVIS checkpoints carry a temp parameter, but the conversion script drops it. convert_blip_2_original_to_pytorch.py L223 asserts it away as an unexpected key, so converted models have no temperature to scale similarities with. This PR adds logit_scale as a learnable parameter on the class, constant-initialised from a new Blip2Config.logit_scale_init_value (2.6592, matching CLIP). It does not read the LAVIS temp value; mapping that through conversion would be a reasonable follow-up.

2. The loss itself. A new return_loss argument on forward computes symmetric cross entropy over logits_per_image and logits_per_text, each scaled by logit_scale.exp(). The two directions are averaged and returned in Blip2ImageTextMatchingModelOutput.loss (and in the tuple path when return_dict=False).

Backwards compatible. return_loss defaults to False, and the returned logits stay unscaled either way, so existing inference behaviour is unchanged. There's a test for this.

Scope. ITC only. return_loss=True with use_image_text_matching_head=True raises a ValueError rather than silently doing the wrong thing. This is deliberately not the full QFormer pretraining recipe from #34019: no hard-negative ITM sampling, no ITG. @qgallouedec noted that's a training recipe rather than a forward(labels=...) addition. Partial progress on #34019.

Test plan

New tests in tests/models/blip_2/test_modeling_blip_2.py:

• loss is None when return_loss is not passed
• return_loss=True gives a finite scalar loss
• logits are identical with and without the loss computation
• backward reaches logit_scale.grad
• ITM head + return_loss=True raises ValueError

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 and the Pull Request checks?
• Was this discussed/approved via a Github issue or the forum? Please add a link to it if that's the case. — Add Loss Functions for QFormer Training in BLIP-2 Model (ITC, ITM, and ITG) #34019
• Did you make sure to update the documentation with your changes according to the guidelines?
• Did you write any new necessary tests?

Who can review?

@qgallouedec — you offered to look at a scoped draft on this class in #34019.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: blip_2

[github-actions]

---

CI recap

Dashboard: View test results in Grafana
Latest run: 34065895245
Result: success | Grafana metrics are not available yet.