[Feature] Add SigLIP loss and forward template

[24GUNV]

Description

Closes #284

This pull request adds SigLIP support for image-text pretraining as described in Zhai et al., 2023.

I referenced the SigLIP loss implementations from Google Research Big Vision and OpenCLIP.

Changes

• Added SigLIPLoss class in stable_pretraining/losses/multimodal.py.
• Implements the pairwise sigmoid image-text objective with learnable logit scale and logit
  bias.
• Supports distributed feature gathering so DDP training uses cross-rank negatives.

Added forward template

• Added siglip_forward in stable_pretraining/forward.py.
• Supports paired image/text batches with vision_model, text_model, and siglip_loss.
• Accepts model outputs exposing either image_embeds / text_embeds or pooler_output.

Added example

• Created benchmarks/diffusiondb2m/siglip-vitb.py.
• Uses Hugging Face SigLIP vision/text towers with the reusable forward template.

Added tests

• Created stable_pretraining/tests/unit/test_siglip_loss.py.
• Added siglip_forward coverage in stable_pretraining/tests/unit/test_forward_functions.py.
• Tests cover SigLIP loss behavior, reference formula, normalization, learnable parameter
  gradients, unpaired batch validation, and forward template outputs.

Docs

• Added API docs for SigLIPLoss and siglip_forward.
• Added the SigLIP citation to docs/source/references.bib.

Checklist

• I have read the Contributing document.
• The documentation is up-to-date with the changes I made (check build artifacts).
• All tests passed, and additional code has been covered with new tests.
• I have added the PR to the RELEASES.rst file.