Configure fixed sampling sigmas on QwenImage21Pipeline

[naykun]

What does this PR do?

Follow up on the review in #3 by storing sample_sigmas in the QwenImage21 pipeline config (model_index.json) instead of extending FlowMatchEulerDiscreteScheduler. The scheduler implementation is unchanged.

The pipeline passes the selected grid through the scheduler's existing sigmas argument. For a fixed distilled grid, the checkpoint scheduler config disables dynamic shifting, terminal stretching and sigma conversions, with shift=1.0. The docs include the checkpoint configuration and save/load workflow.

Explicit call-time sigmas override the saved pipeline grid. Otherwise the configured grid determines the step count; num_inference_steps is used only when neither grid is supplied. Loading with from_pretrained(..., sample_sigmas=...) can also override the saved default. Explicit grids still go through the scheduler's configured processing.

Validation

• Four CPU regression cases passed: configured-grid step counts, explicit overrides, and pipeline save/load including a load-time override.
• Real accelerated checkpoint: 2048 × 2048, 8 steps, bf16, one fixed text prompt and seed on a single 80 GiB GPU. The new pipeline-config path and the previous scheduler-config implementation produced exactly equal sigmas, timesteps, every denoising latent, and final RGBA pixels (maximum difference 0). Both paths shared the same model components to isolate the scheduling change.
• make style and make fix-copies completed. The final three changed files match the snapshot used in the real-checkpoint test.

This verifies parity for that fixed grid and input. Quality with other grids or four-step sampling has not been evaluated; passing num_inference_steps=4 alone still uses the configured eight-point grid.

Self-review

READY. No blocking issues or unused added code found. The new config is exercised through construction, serialization, loading and inference; regression tests call the real scheduler and pipeline. Parameter precedence and scheduler preprocessing are documented. No unresolved review findings; the validation scope is stated above.