fix(qwen3_5): shard ViT frames across CP ranks

[kcz358]

What

Updates Qwen3.5 ViT frame parallel dispatch so CP ranks are symmetric sources:

• Each CP rank first takes a deterministic contiguous shard of its duplicated local frames.
• The existing LPT balancing still runs across the flat dp×cp group.
• Reverse dispatch returns each CP rank's frame shard.
• CP ranks then use an autograd-aware variable-length all-gather to reconstruct the full local-dp feature sequence before masked scatter.

Why

The previous implementation made cp_rank=0 the only source and cp_rank>0 pure receivers, then used CP all_reduce as a broadcast. That creates source/receiver autograd graph asymmetry. It works for some qwen3_5 video cases but can deadlock in more complex wrappers such as aero_realtime with SP + ViT frame parallel: cp_rank=0 remains in backward while cp_rank>0 reaches grad clipping.

This keeps the same global dp×cp load balancing but removes cp source/receiver asymmetry.

Validation

• Verified syntax with .
• Manually validated in training: aero_realtime with SP + vit_frame_parallel no longer deadlocks; qwen3_5 SP + vit_frame_parallel + video still runs.