fix: deduplicate timesteps in DPMSolverMultistep for squaredcos_cap_v2 + karras/lu sigmas

[Chase-Xuu]

Summary

Fixes #12771 — DPMSolverMultistepScheduler crashes with IndexError: index 21 is out of bounds for dimension 0 with size 21 when using beta_schedule='squaredcos_cap_v2' with use_karras_sigmas=True.

Root Cause

When using karras (or lu) sigmas with the cosine (squaredcos_cap_v2) beta schedule, _sigma_to_t() returns float timestep values that are extremely close together at the high end of the schedule:

[998.999, 998.903, 998.802, 998.696, 998.583, 998.464, 998.338, 998.203, 998.059, ...]

These were intentionally not rounded for the cosine schedule (unlike linear and scaled_linear), but since they're cast to int64 downstream, they all collapse to the same value (998), creating 9 duplicate timesteps.

The duplicate timesteps cause index_for_timestep() to start at index 1 (it picks the second occurrence of duplicates). After 20 step iterations, step_index reaches 21, and the second-order update's self.sigmas[self.step_index + 1] goes out of bounds on a size-21 array.

Fix

1. Always round timesteps for karras/lu sigma schedules — this matches the behavior of DPMSolverSinglestepScheduler, which already rounds unconditionally (line 386)
2. Deduplicate timesteps and their corresponding sigmas after rounding to prevent step index drift

When deduplication occurs, the number of effective inference steps is reduced (e.g., 20 requested → 12 unique). This is correct behavior since the redundant timesteps would have produced identical denoising steps anyway.

Test Results

Configuration	Before	After
squaredcos_cap_v2 + karras (20 steps)	❌ IndexError at step 19	✅ 12 unique steps, completes
linear + karras (20 steps)	✅	✅ Unaffected (no duplicates)
squaredcos_cap_v2 without karras	✅	✅ Unaffected
squaredcos_cap_v2 + lu_lambdas	❌ Same index drift	✅ 17 unique steps, completes

Reproduction

from diffusers import DPMSolverMultistepScheduler
import torch

scheduler = DPMSolverMultistepScheduler(
    beta_schedule='squaredcos_cap_v2',
    use_karras_sigmas=True,
)
scheduler.set_timesteps(20)

latents = torch.randn(1, 4, 64, 64)
noise_pred = torch.randn_like(latents)
for i, t in enumerate(scheduler.timesteps):
    result = scheduler.step(noise_pred, t, latents)
    latents = result.prev_sample
# Before: IndexError at step 19
# After: Completes successfully