Fix previous timestep of DDIMInverseScheduler

[stop1one]

What does this PR do?

Fixes #10695

Reproduction

import torch
import numpy as np
from diffusers import DDIMInverseScheduler

num_train_timesteps=1000
scheduler = DDIMInverseScheduler(num_train_timesteps=num_train_timesteps, timestep_spacing='leading')

inference_step = 10
scheduler.set_timesteps(inference_step)

# Before fix: The previous timestep can become negative.
previous_timesteps = torch.tensor([min(timestep - num_train_timesteps // inference_step, num_train_timesteps - 1) for timestep in scheduler.timesteps])
print('Previous Timestep\tCurrent timesteps')
for prev, cur in zip(previous_timesteps, scheduler.timesteps):
    print(f'{prev}\t\t\t{cur}')

# After fix
previous_timesteps = torch.tensor([min(timestep + num_train_timesteps // inference_step, num_train_timesteps - 1) for timestep in scheduler.timesteps])
print('Previous Timestep\tCurrent timesteps')
for prev, cur in zip(previous_timesteps, scheduler.timesteps):
    print(f'{prev}\t\t\t{cur}')

Just changed the sign when calculating the previous timestep.

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

@yiyixuxu