Add FLUX model to Hub

[maitry63]

This PR fixes FLUX model to KerasHub, including the model implementation, checkpoint conversion utilities, and preset generation.

Fixes: #2938

Changes -

• Added the FLUX backbone implementation to KerasHub.
• Added FLUX transformer layers and supporting math utilities.
• Added support for:
  • Double-stream transformer blocks.
  • Single-stream transformer blocks.
  • Timestep embeddings.
  • Positional embeddings.
  • Modulation and AdaLN layers.
  • Attention and MLP projections.
  • Final output projection.
• Updated the FLUX model construction and Functional graph handling.
• Added FLUX configuration conversion from the Hugging Face configuration.
• Added FLUX checkpoint weight conversion from the original PyTorch/Safetensors
  checkpoint format.
• Supports conversion of modulation weights for both dedicated modulation
  layers and Sequential-based implementations.
• Added the FLUX.1 Schnell checkpoint-to-KerasHub preset generation script.
• Generated a KerasHub preset from the official FLUX.1 Schnell checkpoint.
• Ensured PyTorch linear weights are transposed correctly when loading them into Keras layers.

Model Configuration -

• Hidden size: 3072
• Attention heads: 24
• Double-stream blocks: 19
• Single-stream blocks: 38
• Input channels: 64
• MLP ratio: 4.0
• Positional embedding axes: [16, 56, 56]
• Guidance embedding: disabled

Validation -

• Successfully constructed the FluxBackbone.
• Successfully converted the FLUX.1 Schnell Safetensors checkpoint.
• Successfully mapped the checkpoint weights into the Keras model.
• Successfully generated the KerasHub FLUX.1 Schnell preset.
• Verified the conversion workflow end-to-end.

Checklist

• I have added all the necessary unit tests for my change.
• I have verified that my change does not break existing code and works with all backends (TensorFlow, JAX, and PyTorch).
• My PR is based on the latest changes of the main branch (if unsure, rebase the code).
• I have followed the Keras Hub Model contribution guidelines in making these changes.
• I have followed the Keras Hub API design guidelines in making these changes.
• I have signed the Contributor License Agreement.

[gemini-code-assist]

Code Review

This pull request refactors the FLUX model implementation in KerasHub, including converting DoubleStreamBlock to a Layer, optimizing tensor reshaping and attention logic for backend-agnostic static shape inference, and adding preset configurations and weight conversion utilities for FLUX.1. Feedback on these changes focuses on replacing placeholder organization names in Kaggle handles, adhering to repository style guides by saving initialization arguments as attributes and implementing get_config() for DoubleStreamBlock, removing redundant monkey-patching in the checkpoint conversion script, and correcting default input shapes in FluxBackbone to match the actual FLUX architecture.

[gemini-code-assist]

The default shapes for image_shape, text_shape, and y_shape do not match the actual dimensions used by the FLUX model (which has 64 channels for latent images, 4096 channels for T5 text embeddings, and 768 dimensions for the pooled vector y). If a user instantiates the model with default arguments, they will encounter shape mismatch errors when feeding standard FLUX inputs. Please update the defaults to match the actual FLUX architecture.

Suggested change

	image_shape=(None, 3072),
	text_shape=(None, 3072),
	image_ids_shape=(None, 3),
	text_ids_shape=(None, 3),
	y_shape=(128,),
	image_shape=(None, 64),
	text_shape=(None, 4096),
	image_ids_shape=(None, 3),
	text_ids_shape=(None, 3),
	y_shape=(768,),