Added minimal TransformerBridge support for Qwen2MoeForCausalLM

[Canonik]

Description

Added minimal TransformerBridge support for "Qwen2MoeForCausalLM".

This adapter reuses the existing Qwen2 attention path and existing MoE bridge components, while adding the Qwen2-MoE-specific shared expert path.

The PR includes :

• registering/detecting "Qwen2MoeForCausalLM"
• mapping Qwen2-MoE config fields from HF
• Qwen2-style attention/GQA plumbing
• routed expert MoE path
• shared expert MLP path
• shared expert gate
• local tiny config HF parity tests
• cache coverage for MoE related hook points

Fixes #1464

Type of change

• New model / architecture support
• Tests

Testing

Ran the narrow Qwen2-MoE checks:

UV_NO_SYNC=1 uv run pytest tests/unit/model_bridge/supported_architectures/test_qwen2_moe_adapter.py
UV_NO_SYNC=1 uv run pytest tests/integration/model_bridge/test_qwen2_moe_bridge.py
UV_NO_SYNC=1 uv run pytest tests/unit/tools/test_model_registry.py -k TestRegistrySyncedWithFactory

[Canonik]

should be good to go

[jlarson4]

Great work on this @Canonik! Thank you for getting this taken care of, I am merging now