Fix CogVideoX dynamic CFG to use loop index instead of timestep value

[veeceey]

The dynamic CFG calculation in CogVideoX pipelines was using t.item() (the raw timestep value, which ranges from ~1 to ~999) instead of the loop index i (which ranges from 0 to num_inference_steps - 1). This made ((num_inference_steps - t.item()) / num_inference_steps) go deeply negative, and after ** 5.0 + math.cos, the guidance scale oscillates wildly instead of following a smooth cosine annealing schedule.

Replaced t.item() with i in all four CogVideoX pipeline files:

• pipeline_cogvideox.py
• pipeline_cogvideox_image2video.py
• pipeline_cogvideox_video2video.py
• pipeline_cogvideox_fun_control.py

Quick sanity check with guidance_scale=4, num_inference_steps=50:

OLD (broken): 1.86, 2.03, 4.98, 4.61, 2.19, 3.18, 1.03, 3.11, 1.41, 4.99 ...
NEW (fixed):  5.00, 4.91, 4.67, 4.34, 3.96, 3.56, 3.17, 2.81, 2.49, 2.21 ...

Fixes #9641

[veeceey]

Manual Test Results

Verified the dynamic CFG schedule with guidance_scale=4, num_inference_steps=50:

Before fix (oscillating wildly due to t.item() being 1-999):

Step 0: t=999, cfg=1.8610
Step 1: t=979, cfg=2.0290
Step 2: t=959, cfg=4.9829
Step 3: t=939, cfg=4.6136
Step 4: t=919, cfg=2.1921
Step 5: t=899, cfg=3.1813
Step 6: t=879, cfg=1.0251
Step 7: t=859, cfg=3.1136
Step 8: t=839, cfg=1.4120
Step 9: t=819, cfg=4.9888

After fix (smooth cosine annealing from 5.0 down to 1.0):

Step 0: t=999, cfg=5.0000
Step 1: t=979, cfg=4.9096
Step 2: t=959, cfg=4.6729
Step 3: t=939, cfg=4.3409
Step 4: t=919, cfg=3.9584
Step 5: t=899, cfg=3.5609
Step 6: t=879, cfg=3.1740
Step 7: t=859, cfg=2.8145
Step 8: t=839, cfg=2.4917
Step 9: t=819, cfg=2.2100

Also fixed in pipeline_cogvideox_fun_control.py which had the same bug.