Fix NaN corruption on final step in DPMSolverMultistep and UniPCMultistep with final_sigmas_type="sigma_min"

[deep1blue]

What does this PR do?

Fixes #14890.

Root Cause

When final_sigmas_type="sigma_min" is enabled alongside sigma schedules whose final step already reaches sigma_min (use_karras_sigmas=True, use_exponential_sigmas=True, use_beta_sigmas=True, or use_lu_lambdas=True), set_timesteps() appends sigma_last = sigmas[-1] to the sigma array, making the last two entries identical (sigmas[-2] == sigmas[-1]).

On the final denoising step (step_index == num_inference_steps - 1), this produces a zero-width log-SNR step (h = lambda_t - lambda_s == 0.0). While dpm_solver_first_order_update safely multiplies by expm1(-h) = 0.0 (leaving the sample unchanged), the 2nd-order and 3rd-order multistep updates in DPMSolverMultistepScheduler (and the multistep predictor/corrector in UniPCMultistepScheduler) divide h = 0.0 by r0 = h_0 / h = inf (or compute hh / h_0 where h_0 = 0.0), producing 0.0 / 0.0 = NaN and corrupting the entire output tensor on the very last step.

Solution

• In DPMSolverMultistepScheduler.step(): include (self.step_index == self.num_inference_steps - 1 and self.sigmas[self.step_index + 1] == self.sigmas[self.step_index]) in lower_order_final so zero-width terminal steps use dpm_solver_first_order_update.
• In UniPCMultistepScheduler.step(): bypass the multistep corrector when the previous step had zero width (self.sigmas[self.step_index] == self.sigmas[self.step_index - 1]) and fall back to first-order multistep_uni_p_bh_update(..., order=1) when the final step has zero width (self.sigmas[self.step_index + 1] == self.sigmas[self.step_index]).
• Added unit tests test_zero_width_final_step in both tests/schedulers/test_scheduler_dpm_multi.py and tests/schedulers/test_scheduler_unipc.py.

[deep1blue]

Gentle, polite ping @stevhliu @sayakpaul @yiyixuxu @DN6 @hlky — whenever you have a spare moment, would you mind taking a quick look at this small, self-contained scheduler fix for #14890? When final_sigmas_type="sigma_min" duplicates the terminal sigma (sigmas[-2] == sigmas[-1], h == 0), falling back to the exact first-order update on the zero-width terminal step avoids the 0.0 / 0.0 = NaN division in DPMSolverMultistepScheduler and UniPCMultistepScheduler, verified with unit tests across both schedulers. Thank you so much for your time!