feat: Add Qwen3.5 support to Megatron backend

[vivekkalyan]

Summary

This adds Qwen3.5 MoE support to the Megatron backend on top of #636.

With this stack, ART can now train Qwen/Qwen3.5-35B-A3B with Megatron while serving inference from the dedicated vLLM process using merged weight updates.

What changed

• Bump megatron-bridge to a revision that includes upstream Qwen3 support needed for this path
• Teach the Megatron provider path to accept Qwen3.5 MoE Bridge models and patch the hybrid layer spec so ART's flex attention only applies to the standard attention layers
• Add Qwen3.5 LoRA coverage for:
  • gated delta net in-projection
  • shared experts
  • the Qwen3.5 attention packing layout while preserving existing Qwen3 behavior
• Replace the old ART-specific merged-weight math with a Bridge compatibility layer that converts ART adapters into Bridge AdapterWeights for export and merge
• Add provider, wrapper, and helper coverage for the Qwen3.5 path

Validation

• Clean-cluster Qwen3.5 Megatron dedicated + merged smoke completed through step @2
• Clean-cluster 20-step yes/no/maybe run completed successfully for Qwen/Qwen3.5-35B-A3B