Fix division by zero in rescale_noise_cfg causing NaNs during inference

[Akash504-ai]

What does this PR do?

This PR fixes a numerical stability issue in rescale_noise_cfg that can lead to division by zero and produce NaN/inf values during inference.

Problem

The current implementation performs:

noise_pred_rescaled = noise_cfg * (std_text / std_cfg)

If std_cfg becomes zero (e.g., when noise_cfg has zero variance), this results in division by zero, producing NaNs that can silently corrupt the diffusion process.

Solution

A small epsilon value is added to the denominator to ensure numerical stability:

noise_pred_rescaled = noise_cfg * (std_text / (std_cfg + 1e-6))

Scope

This fix is applied consistently across:

• src/diffusers/guiders/guider_utils.py
• src/diffusers/pipelines/stable_diffusion/pipeline_stable_diffusion.py
• src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py

Impact

• Prevents NaN/inf propagation during inference
• Improves numerical stability in edge cases
• No impact on normal behavior when std_cfg > 0

Fixes #13425

@sayakpaul @yiyixuxu