Add diverse examples for Cache-dit in Flux.1 and Qwen models regarding #763

[Bluear7878]

This PR adds comprehensive example scripts for various Flux.1 and Qwen-Image models to verify and utilize the #763.

I have reviewed and tested PR #763, confirming it works as expected. To further demonstrate its capabilities and ensure broad compatibility, I have added example scripts for the following models:

• Flux.1 Series: Canny, Depth, Fill, Kontext, Krea, Redux, and Dev-cache
• Qwen-Image Series: Base, Edit, and ControlNet

Benchmark Results

I evaluated the performance by comparing the inference time between No_cache and cache_dit. The experiments were conducted on an RTX PRO 6000 Blackwell Max-Q.

The results demonstrate significant latency improvements across most models, with Flux.1 Fill achieving up to 1.92x speedup.

Model	Steps	No Cache (s)	Cache DiT (s)	Speedup
flux.1-canny-dev	20	4.49	3.75	1.20x
flux.1-depth-dev	20	4.73	4.26	1.11x
flux.1-fill-dev	30	9.90	5.16	1.92x
flux.1-kontext-dev	20	9.33	7.00	1.33x
flux.1-krea-dev	20	4.34	3.41	1.27x
flux.1-redux-dev	20	4.56	4.03	1.13x
flux.1-dev-cache (vs DoubleFBCache)	50	4.52	5.34	0.85x
qwen-image-edit-2509	40	148.23	115.13	1.29x
qwen-image	50	54.01	43.09	1.25x
qwen-image-controlnet	30	35.82	32.44	1.10x