[FEATURE] Add native Qwen-Image 2.1 support

[lzx1413]

Description

Integrate Qwen-Image 2.1 into TeleFuser with a locally implemented DiT and a stage-based pipeline for text-to-image generation, image editing, and reference-guided generation.

Motivation

Enable Qwen-Image 2.1 inference through TeleFuser’s existing pipeline and model-loading architecture, with runnable examples and service contract coverage.

Type of Change

• New feature (non-breaking change which adds functionality)
• Documentation update

Changes Made

• Implement the Qwen-Image 2.1 DiT and compose text encoding, denoising, and VAE stages using ModuleManager.
• Add text-to-image, edit, and reference examples with defaults declared directly in Python. The reference example uses the official five-image “Outfit styling” case.
• Update the example README and add example registry and service parity tests.

Testing

• Focused unit tests pass: 69 tests
• Manual reference generation performed on physical GPU 3
• Benchmarks added or updated

pytest -q tests/unit/service/test_example_service_parity.py tests/unit/test_example_registry.py
ruff check examples/qwen_image telefuser/models/qwen_image_21_dit.py telefuser/pipelines/qwen_image tests/unit/service/test_example_service_parity.py tests/unit/test_example_registry.py
CUDA_VISIBLE_DEVICES=3 python examples/qwen_image/qwen_image_21_reference_h100.py \
  --output_path work_dirs/qwen-image-2.1-reference.png

The 40-step reference run produced a valid 1024×1024 image. Python syntax and diff checks passed.

Checklist

• Code follows the project’s coding standards; ruff and ruff-format commit hooks passed
• Full pre-commit suite run (pre-commit run --all-files)
• Full test suite passes (pytest tests/)
• New focused tests added
• Documentation updated
• Commit message is clear and descriptive
• PR title follows the [TYPE] Brief description convention

Related Issues

None linked.

Additional Notes

The existing service schema accepts one reference image per request. The five-image official example runs through the CLI without adding a new service interface.

GPU Architecture Support

No kernels were added or modified. Inference was validated on an H100 (SM90).

Performance Impact

No performance benchmark was run; this PR makes no performance claim.