[Vuln3rat0r]

Summary

Fixes #15167.

When use_flow_sigmas=True, multistep schedulers previously cast self.timesteps to torch.int64, truncating continuous flow-matching timesteps into integer values. Continuous flow models (such as FLUX.1, SD3.5, Wan2.1, and Cosmos) require floating-point timesteps to integrate along continuous trajectories without truncation artifacts.

This change preserves torch.float32 timesteps when self.config.use_flow_sigmas=True, and preserves torch.int64 for discrete diffusion schedulers (use_flow_sigmas=False).

Schedulers updated:

• UniPCMultistepScheduler
• DPMSolverMultistepScheduler
• DPMSolverSinglestepScheduler
• DPMSolverMultistepInverseScheduler
• DEISMultistepScheduler
• SASolverScheduler

Testing

• Added test_flow_sigmas_continuous_timesteps across all updated scheduler test suites to verify torch.float32 timestep retention and end-to-end integration step execution.
• Updated Cosmos 3 pipeline test assertions to match continuous timestep values.
• Verified all scheduler test suites and code formatting passes (ruff check, ruff format --check).

Reviewers

@yiyixuxu @dg845