performance(ring para): change sdpa backend to flash attention

[JasonHe-WQ]

Description

When using the sdpa attention backend with ring sp, this pr replaces efficient backend with flash backend to accelerate.

Related Issue

#266

Test Env

config	val
Python / PyTorch	3.12.13 / 2.11.0+cu130
CUDA / NCCL	13.0 / 2.28.9
Diffusers / Transformers	0.36.0 / 4.57.6
GPU	4 x NVIDIA RTX PRO 5000 72GB Blackwell，driver 580.159.03
Attention backend	DiffSynth sdpa
Dtype	BF16
Picture	1024 x 1024，1 per prompt
Denoise	5 steps，CFG scale 4.0，seed 42
Prompt	A red fox sitting beside a mountain lake at sunrise, highly detailed
Negative prompt	blurry, low quality

Precision

	Cosine	Normalized SSIM	PSNR	MAE / 255	Max abs
flash CP2 vs efficient CP2	0.997295	0.969289	28.049 dB	4.750 / 255	196
flash CP2 vs SP2	0.997394	0.969862	28.226 dB	4.501 / 255	187
efficient CP2 vs SP2	0.997398	0.970382	28.218 dB	4.571 / 255	180

From tables above, we can conclude that comparing to efficient CP2, flash ring CP2 achieves comparable precision to ulysses SP2.

Throughput

mode	Request wall time	Denoise stage	steady step
Flash CP2	2.946104 +/- 0.010519 s	2750.312 +/- 12.587 ms	550.423 +/- 2.351 ms
SP2	3.145325 +/- 0.011289 s	2951.521 +/- 12.741 ms	590.334 +/- 2.524 ms
Efficient CP2	3.459390 +/- 0.025342 s	3260.770 +/- 19.557 ms	652.187 +/- 4.197 ms

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.