Fix Flux IP-Adapter masks, CFG embeds, and modular helper coupling

[taivu1998]

What this PR does

Addresses #13584 by fixing several Flux pipeline and model issues found in the Flux review:

• Applies ip_adapter_masks in FluxIPAdapterAttnProcessor instead of accepting and silently ignoring them.
• Enables true CFG when users pass precomputed negative_prompt_embeds and negative_pooled_prompt_embeds to Flux img2img, inpaint, and ControlNet pipelines.
• Validates positive and negative prompt embedding shape mismatches in base Flux and Flux Kontext.
• Fixes Flux ControlNet img2img dimension validation so it checks divisibility by vae_scale_factor * 2.
• Hardens Flux Prior Redux batch validation for tensor/NumPy image batches, prompt lists, prompt embeddings, and both prompt scale lists.
• Removes hidden Flux modular pipeline dependencies on classic Flux pipeline private helpers and QwenImage helpers by adding Flux-local modular helper utilities.

Root cause

The affected Flux code paths had drifted across pipeline variants:

• Flux attention already accepted ip_adapter_masks, but the processor never consumed them.
• Some true-CFG-capable pipelines only checked negative_prompt, even though their input validation allowed negative embeddings.
• Base Flux/Kontext validation missed a shape check that sibling Flux pipelines already had.
• ControlNet img2img used height % self.vae_scale_factor * 2, which Python parses as (height % self.vae_scale_factor) * 2.
• Prior Redux only counted list images when validating batch-coupled prompt and scale inputs.
• Modular Flux reused private helpers from the classic Flux pipeline and a helper from QwenImage, creating cross-family coupling.

Validation

Ran the focused regression suite for the changed behavior:

uvx ruff format src/diffusers/models/transformers/transformer_flux.py src/diffusers/modular_pipelines/flux/before_denoise.py src/diffusers/modular_pipelines/flux/decoders.py src/diffusers/modular_pipelines/flux/encoders.py src/diffusers/modular_pipelines/flux/inputs.py src/diffusers/modular_pipelines/flux/pipeline_helpers.py src/diffusers/pipelines/flux/pipeline_flux.py src/diffusers/pipelines/flux/pipeline_flux_controlnet.py src/diffusers/pipelines/flux/pipeline_flux_controlnet_image_to_image.py src/diffusers/pipelines/flux/pipeline_flux_img2img.py src/diffusers/pipelines/flux/pipeline_flux_inpaint.py src/diffusers/pipelines/flux/pipeline_flux_kontext.py src/diffusers/pipelines/flux/pipeline_flux_prior_redux.py tests/models/transformers/test_models_transformer_flux.py tests/modular_pipelines/flux/test_modular_pipeline_flux.py tests/pipelines/controlnet_flux/test_controlnet_flux.py tests/pipelines/controlnet_flux/test_controlnet_flux_img2img.py tests/pipelines/flux/test_pipeline_flux.py tests/pipelines/flux/test_pipeline_flux_img2img.py tests/pipelines/flux/test_pipeline_flux_inpaint.py tests/pipelines/flux/test_pipeline_flux_kontext.py tests/pipelines/flux/test_pipeline_flux_redux.py
uvx ruff check src/diffusers/models/transformers/transformer_flux.py src/diffusers/modular_pipelines/flux/before_denoise.py src/diffusers/modular_pipelines/flux/decoders.py src/diffusers/modular_pipelines/flux/encoders.py src/diffusers/modular_pipelines/flux/inputs.py src/diffusers/modular_pipelines/flux/pipeline_helpers.py src/diffusers/pipelines/flux/pipeline_flux.py src/diffusers/pipelines/flux/pipeline_flux_controlnet.py src/diffusers/pipelines/flux/pipeline_flux_controlnet_image_to_image.py src/diffusers/pipelines/flux/pipeline_flux_img2img.py src/diffusers/pipelines/flux/pipeline_flux_inpaint.py src/diffusers/pipelines/flux/pipeline_flux_kontext.py src/diffusers/pipelines/flux/pipeline_flux_prior_redux.py tests/models/transformers/test_models_transformer_flux.py tests/modular_pipelines/flux/test_modular_pipeline_flux.py tests/pipelines/controlnet_flux/test_controlnet_flux.py tests/pipelines/controlnet_flux/test_controlnet_flux_img2img.py tests/pipelines/flux/test_pipeline_flux.py tests/pipelines/flux/test_pipeline_flux_img2img.py tests/pipelines/flux/test_pipeline_flux_inpaint.py tests/pipelines/flux/test_pipeline_flux_kontext.py tests/pipelines/flux/test_pipeline_flux_redux.py
git diff --check
uv run --python 3.10 python -m py_compile src/diffusers/models/transformers/transformer_flux.py src/diffusers/modular_pipelines/flux/before_denoise.py src/diffusers/modular_pipelines/flux/decoders.py src/diffusers/modular_pipelines/flux/encoders.py src/diffusers/modular_pipelines/flux/inputs.py src/diffusers/modular_pipelines/flux/pipeline_helpers.py src/diffusers/pipelines/flux/pipeline_flux.py src/diffusers/pipelines/flux/pipeline_flux_controlnet.py src/diffusers/pipelines/flux/pipeline_flux_controlnet_image_to_image.py src/diffusers/pipelines/flux/pipeline_flux_img2img.py src/diffusers/pipelines/flux/pipeline_flux_inpaint.py src/diffusers/pipelines/flux/pipeline_flux_kontext.py src/diffusers/pipelines/flux/pipeline_flux_prior_redux.py tests/models/transformers/test_models_transformer_flux.py tests/modular_pipelines/flux/test_modular_pipeline_flux.py tests/pipelines/controlnet_flux/test_controlnet_flux.py tests/pipelines/controlnet_flux/test_controlnet_flux_img2img.py tests/pipelines/flux/test_pipeline_flux.py tests/pipelines/flux/test_pipeline_flux_img2img.py tests/pipelines/flux/test_pipeline_flux_inpaint.py tests/pipelines/flux/test_pipeline_flux_kontext.py tests/pipelines/flux/test_pipeline_flux_redux.py
uv run --python 3.10 python -m pytest tests/models/transformers/test_models_transformer_flux.py tests/pipelines/flux/test_pipeline_flux_redux.py tests/modular_pipelines/flux/test_modular_pipeline_flux.py tests/pipelines/flux/test_pipeline_flux.py tests/pipelines/flux/test_pipeline_flux_kontext.py tests/pipelines/controlnet_flux/test_controlnet_flux_img2img.py tests/pipelines/flux/test_pipeline_flux_img2img.py tests/pipelines/flux/test_pipeline_flux_inpaint.py tests/pipelines/controlnet_flux/test_controlnet_flux.py -k "ip_adapter_masks_are_applied or ip_adapter_masks_validate_num_images or ip_adapter_no_mask_keeps_3d_hidden_state_support or FluxReduxFastTests or modular_blocks_do_not_import or latent_helpers_roundtrip or negative_prompt_embeds_shape_mismatch or dimension_check_uses_packed_latent_multiple or true_cfg_with_negative_prompt_embeds"

Targeted pytest result:

15 passed, 497 deselected

Notes

This PR adds focused fast regression coverage for the fixed behavior. The broader slow/nightly real-checkpoint coverage discussed in the issue is intentionally left for a separate CI-focused PR because it spans many Flux variants and checkpoint/runtime choices.