Add MoVA inference and MP2 parity support

[akhauriyash]

Summary

• add the MoVA inference layer and wire it into the xLLM model
• support the model-parallel-2 MoVA routing and weight layout used by the RL stack
• preserve required top-k routing behavior
• add focused MoVA, transfer-engine, and CPU top-k coverage

Stack dependencies

This is the SGLang side of the internal MoVA RL stack and is reviewed together with:

• Miles counterpart: Add MoVA training, weight sync, and rollout validation miles#87
• RL360 integration: https://github.com/LLM360/RL360/pull/492
• Megatron counterpart: LLM360/Megatron-LM@prod...codex/mova-post-fc2-prob-20260814

The validated training recipe must use --no-rope-fusion for MoVA. The pinned Megatron fused packed-THD RoPE path does not preserve rotary_interleaved=True, causing trainer/rollout policy mismatch. Keep fusion disabled until the permanent Megatron fix passes end-to-end parity.

Validation

• exact head: 198d611c005b789879ac73732e67d0d55d6362fd
• base/head comparison: prod...yash.akhauri/xllm-router-mp2-parity
• full 16K, global-batch-64 integration run reached 89 optimizer updates before an unrelated checkpoint-write failure: https://wandb.ai/mbzuai-llm/mova-dapo-math-rl/runs/so5ey8aq
• completed smoke run: https://wandb.ai/mbzuai-llm/mova-dapo-math-rl/runs/p4jl7aqs
• branch adds registered tests for MoVA layer behavior and remote weight transfer, plus CPU top-k coverage

Draft for internal review. No merge requested.