Optimize diffusion calibration step sampling

[changwangss]

Description

This PR reduces diffusion calibration cost by decoupling calibration denoising steps from the full inference schedule.

Main changes:

• Add calib_num_inference_steps, defaulting to 8, for diffusion calibration.
• Pass calib_num_inference_steps directly to the diffusion pipeline, allowing the scheduler to construct its native short schedule.
• Support higher-order schedulers without manually slicing or modifying scheduler timesteps/sigmas.
• Forward custom pipeline_call_kwargs during calibration.
• Preserve calibration-owned values for guidance_scale, generator, and num_inference_steps.
• Propagate pipeline exceptions, including NotImplementedError, instead of silently treating failed calls as valid samples.
• Keep diffusion SignRound calibration at micro-batch size 1 and scale gradient accumulation to preserve the effective batch size.
• Update dual-transformer calibration to use calib_num_inference_steps.
• Add CLI and entry-point support for --calib_num_inference_steps.
• Update English and Chinese documentation.

Type of Change

Bug fix / performance improvement

Related Issues

Fixes or relates to #

Checklist Before Submitting

• My code has been tested locally.
• Documentation has been updated as needed.
• New or updated tests are included where applicable.
• The CUDA CI has passed. You can trigger it by commenting /azp run Unit-Test-CUDA-AutoRound.

[AutoRoundBot]

/azp run Unit-Test-CUDA-AutoRound

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[AutoRoundBot]

/azp run Unit-Test-CUDA-AutoRound

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[AutoRoundBot]

/azp run Unit-Test-CUDA-AutoRound

[azure-pipelines]

Azure Pipelines successfully started running 1 pipeline(s).

[lvliang-intel]

LGTM