[Kandinsky 5] Fix I2V conditioning: don't inject the image latent into visual_cond channels

[sashakunitsyn]

What does this PR do?

Kandinsky5I2VPipeline.prepare_latents injects the conditioning image twice: once as the first latent frame (latents[:, 0:1] = image_latents, correct) and again into the DiT's visual_cond conditioning channels (visual_cond[:, 0:1] = image_latents). This duplicate signal over-conditions the first frame, producing mesh/visual artifacts on checkpoints trained with the reference convention. The reference implementation (kandinskylab/kandinsky-5 → kandinsky/generation_utils.py::generate) does not do the second injection — visual_cond is created as zeros. This PR removes that extra injection to match the reference.

Before fix

base_i2v_stock_gh.mp4

After fix

base_i2v_condfix_gh.mp4

[leffff]

@sayakpaul @yiyixuxu
We are preparing for a release of something interesting on top of our Kandinsky5. We found a bug in our I2V implementation. We want to merge this one-line fix ASAP, so we can proceed to our release!
We will be grateful for your review!

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[github-actions]

Hi @sashakunitsyn, thanks for the PR! It does not appear to link an issue it fixes. If this PR addresses an existing issue, please add a closing keyword (e.g. Fixes #1234) to the PR description so the issue is linked. See the contribution guide for more details. If this PR intentionally does not fix a tracked issue, a maintainer can add the no-issue-needed label to silence this reminder.