feat: Add merged vLLM rollout weights

[vivekkalyan]

Enable ART to serve merged LoRA weights through dedicated vLLM so Qwen3.5-MoE training works on the current vLLM build.

Changes

• Add rollout_weights_mode: "lora" | "merged" with dedicated-only validation and a merged-mode requirement for Qwen/Qwen3.5-35B-A3B and Qwen/Qwen3.5-397B-A17B
• Push merged weights into dedicated vLLM with native weight transfer while keeping LoRA checkpoints for training and persistence
• Update dedicated server wiring, validation, and Qwen3.5 smoke scripts/tests for the merged-inference path

[angkywilliam]

Are we able to support training for 397B-A17B?

[vivekkalyan]

not tested for now, but it should work with megatron. we need to add the merging logic to our megatron service as well

[angkywilliam]

Would this be more intuitive if we perform the merging on _merged_checkpoint_weights?

[vivekkalyan]

i prefer to leave _merged_checkpoint_weights to only collect the checkpoint-format weights and normalize the names into the surface vLLM expects.

this keeps the pausing, merging, send weights, unmerge, resume part of the same flow in _sync_merged_weights and would be clearer to understand the flow

i renamed the function to _merged_checkpoint_weights_for_vllm to make it a little clearer that its just a transformation function

[angkywilliam]

We'll probably need to configure separate target_modules for Qwen3.5 Moe model.
This is the config I’ve been using when training Qwen3.5, but it’s worth double checking on the latest Unsloth version.

target_modules=[
            # Full attention layers (25% of layers)
            "q_proj", "k_proj", "v_proj", "o_proj",
            # DeltaNet linear attention layers (75% of layers)
            "in_proj_qkv", "in_proj_z", "in_proj_b", "in_proj_a", "out_proj",
            # MLP (all layers)
            "gate_proj", "up_proj", "down_proj",
            # MoE shared expert gate (only present in MoE models, ignored for dense)
            "shared_expert_gate",
        ],