[Qwen3.5 moe] : fix vision shape mismatch

[quic-mamta]

This PR fixes Qwen3.5 MoE vision shape handling for batched inputs.

Key changes:

• Expands vision rotary positional embeddings across the batch dimension so they align with batched patch tokens.
• Updates dummy Qwen3.5 MoE vision input shapes so pixel_values scales with batch size instead of using a fixed single-grid size.
• Updates the Qwen3.5 MoE disaggregated mode example to:
  • Compile/run vision with a single prompt slot.
  • Filter and cast inputs according to session bindings.
  • Use KV DMA handoff between prefill and decode sessions.
  • Cleanly deactivate and unload sessions.
• Adds CPU unit coverage to verify Qwen3.5 MoE vision output shape for batched patch tokens.