What does this PR do?

Accelerated QwenImage21 checkpoints need a fixed sampling sigma schedule that loads with the model. This PR adds sample_sigmas to the QwenImage21 pipeline config (model_index.json) and passes the selected schedule to the scheduler through its existing sigmas argument.

Model authors configure the default sample_sigmas; users can experiment with an explicit call-time sigmas override. The selected grid determines the step count. When neither grid is supplied, the existing num_inference_steps behavior is preserved. The scheduler implementation is unchanged, and explicit grids still undergo its configured processing.

This incorporates the pipeline-config approach reviewed in #4 and replaces the scheduler changes proposed in #3. Documentation focuses on loading a preconfigured checkpoint and overriding sigmas at runtime.

Validation

• Rebased onto upstream main at c2798cc7859f258c6cfc5b2460e82b6cac71f235; both commit patches are unchanged by the rebase.
• Four CPU regression cases passed after rebase, covering configured-grid step counts, explicit overrides, and pipeline save/load including a load-time override.
• Ruff lint/format, copy consistency, and whitespace checks passed.
• Earlier real-checkpoint validation: 2048 × 2048, 8 steps, bf16, one fixed text prompt and seed. With scheduler transformations disabled for the fixed grid, the pipeline-config and previous scheduler-config implementations produced exactly equal sigmas, timesteps, every denoising latent, and final RGBA pixels (maximum difference 0). Both paths shared model components to isolate the scheduling change; executable changes in this PR are unchanged since that validation.

Real-model parity was checked for that fixed grid and input. Quality with other grids or four-step sampling has not been evaluated; num_inference_steps=4 alone still uses a configured eight-point grid.