Add Flux diffusion model support

[ChuanLi1101]

Summary

Add support for Flux diffusion transformer model for image generation.

Changes

File	Description
atom/models/flux.py	Flux DiT transformer architecture
atom/models/flux_vae.py	VAE encoder/decoder for latent space
atom/models/flux_text_encoder.py	T5/CLIP text encoder wrapper
atom/model_ops/diffusion_sampler.py	Flow matching sampler (Euler)
atom/model_engine/diffusion_runner.py	Diffusion inference pipeline
atom/entrypoints/image_server.py	OpenAI Images API server
atom/examples/simple_image_gen.py	CLI example for image generation
atom/tests/test_flux.py	Unit tests
atom/model_engine/model_runner.py	Register FluxTransformer2DModel

Usage

# API server
python -m atom.entrypoints.image_server --model black-forest-labs/FLUX.1-dev

# CLI example  
python -m atom.examples.simple_image_gen --model black-forest-labs/FLUX.1-dev --prompt "A cat"

Test Results

• Black/Ruff code style checks: PASSED
• 4/6 ATOM LLM tests passed (2 failures are CI runner memory issues, not code related)

[ChuanLi1101]

CI Test Results

Code style checks: All passed (Black, Ruff)

ATOM Tests:

• Meta-Llama-3-8B-Instruct (1 GPU)
• DeepSeek-R1-0528 (8 GPU)
• Qwen3-235B-A22B-Instruct-2507-FP8 (8 GPU)
• gpt-oss-120b (2 GPU)
• Llama-3.3-70B-Instruct-MXFP4-Preview (1 GPU)
• gpt-oss-120b (1 GPU)

Note: The 2 failed tests are due to insufficient GPU memory on single-GPU runners:

• \AssertionError: need at least 2621440 KV cache\ (Llama-3.3-70B)
• \AssertionError: need at least 589824 KV cache\ (gpt-oss-120b)

These failures are CI infrastructure issues, not related to the Flux diffusion model changes. The same tests pass on main when run on multi-GPU runners with sufficient memory.

[ChuanLi1101]

CI Test Failures - Infrastructure Issue

The two failing tests are not related to the Flux code changes in this PR:

Test	Error
\Llama-3.3-70B-Instruct-MXFP4-Preview\	\AssertionError: need at least 2621440 KV cache\
\gpt-oss-120b\	\AssertionError: need at least 589824 KV cache\

Both tests are running on single GPU machines (\�tom-mi355-1gpu), but these 70B/120B models require more GPU memory than available to allocate the KV cache.

This is a CI infrastructure configuration issue - the large models need to run on multi-GPU runners instead.

Request: Could the maintainers please either:

1. Adjust CI configuration to run these large model tests on appropriate hardware, or
2. Skip/exclude these flaky tests for this PR

The branch has been rebased onto the latest \main. Thank you!