[multimodal] add language_model_only flag for models like qwen3.5

[erictang000]

Add language_model_only flag for multimodal models (Qwen3.5)

Summary

• Add language_model_only config flag across policy, ref, and inference engine configs to skip vision encoder initialization for multimodal models like Qwen3.5, reducing GPU memory usage
• Fix FSDP weight sync: remap CausalLM param names (model.layers.*) to vLLM's expected namespace (language_model.model.layers.*) via new weight_prefix in FSDPWeightExtractor
• Make FSDP wrap policy resilient to missing vision-only layer classes (warn + skip instead of crash)
• Add flash-linear-attention and causal-conv1d dependencies; unblock causal-conv1d install override -- required for performant GDN layer execution
• Add run_qwen3.5_0.8b.sh example with use_sample_packing=false (GDN layers are incompatible with packing)

Runs

FSDP and megatron reward matching

Test plan

• Run run_qwen3.5_0.8b.sh on 4 GPUs -- verify weight sync, no GDN fallback warnings, avg_final_rewards trends up
• Run existing non-multimodal FSDP test to confirm no regression
• Verify config validation rejects mismatched language_model_only across policy/ref/generator

---

[devin-ai-integration]

Devin Review found 1 potential issue.

View 5 additional findings in Devin Review.

[erictang000]

cc: @nithinvc PR adding language_model_only flag - this shouldn't effect any of your runs since it's false by default but just heads up