Support Qwen3.5 with FSDP + Megatron Backends

[erictang000]

Requirements for supporting Qwen3.5

vLLM requirements

• transformers==5.3.0 release (needs to include this PR: Fix TypeError in convert_rope_params_to_dict when ignore_keys is a list huggingface/transformers#44272)
• vllm==0.16.1 release for official model release on vLLM (qwen3_5.py)
• a pre-built linux wheel for flash-attn==2.8.3 + torch 2.10.0 (https://github.com/Dao-AILab/flash-attention/releases/tag/v2.8.3 does not include a torch 2.10 wheel, and vllm==0.16.1 requires torch==2.10.0)
• merge the changes to apply_chat_template in this PR: Bump transformers to >=5.0.0 for GLM-4.7-Flash #1241

Megatron requirements:

• megatron-bridge bump commit to main latest: 0034ddaad7fae7c658c3df7e12d13522a4935770

[wxm888888]

How to configure pyproject.toml to perform RL with Qwen 3.5？

[CharlieFRuan]

@wxm888888 This one should work with the dense models (2B, 4B, 9B, 27B) with FSDP: #1313 It is still a WIP PR and pending validating runs, but the pyproject should be configured. Specifically causl-conv1d and flash linear attention are crucial for training speed