Include Qwen shared experts in MoE LoRA

[vivekkalyan]

Summary

• include Qwen3.5/Qwen3.6 MoE shared-expert gate/up/down targets in the default LoRA target set
• wrap and export shared-expert LoRA weights alongside routed grouped experts
• preserve shared-expert target modules when converting fused MoE adapters to vLLM format

Validation

• Sky H200 validation on art-qwen35-shared-expert-lora-test: created a rank-8 Qwen3.5 MoE ART adapter, published it to vLLM format, roundtripped it back into Megatron, compared structure against the run57 Tinker adapter, and loaded it with stock vLLM LoRA loader. Stock vLLM loaded shared-expert gate/up/down and fused routed experts successfully.