Vision CP crashes when a packed sample cannot populate every CP rank

### Reproduction

- Molmo2-4B SFT checkpoint on 8 A100s with CP=8
- `model.apply_cp_to_vision_backbone=true`
- Debug mixture, dynamic packing, 36,864-token context, device batch size 1

### Actual behavior

On the first training step, ranks 5-7 receive image boundaries with identical
start and end indices. The vision backbone passes the resulting zero-image tensor
to SDPA, which fails with:

```text
RuntimeError: batch size must be positive
```

The remaining ranks wait in NCCL until the watchdog terminates the job.
**`compute_image_shard_boundaries()` tries to move whole image/frame boundary units
between ranks, but it cannot make every shard nonempty when there are fewer usable
units than CP ranks.**

Could you kindly provide some help and guidance on this? Thank you! @chrisc36 @sangho-vision @zixianm

Start at compute_image_shard_boundaries() and trace how its image or frame boundary output is consumed by the vision backbone across CP ranks. Reproduce the Molmo2-4B SFT setup from the issue, then verify that ranks do not pass zero-image tensors to SDPA or remain blocked in NCCL when usable units are fewer than CP ranks.

由索引模型根據 Issue 內容生成。