[HunyuanImage3] Add text tasks and tensor parallel inference

[Chernobyllight]

Summary

• extend the native HunyuanImage3 runner with t2t, t2i, ti2t, and ti2i task routing, text-generation inputs, configs, and launchers
• support think_recaption for text-conditioned AR generation and reuse its output in image generation
• add tensor-parallel checkpoint sharding and collectives for attention, dense MLP, MoE, and vocabulary projection
• support both eager and FlashInfer fused MoE under tensor parallel
• add FlashInfer autotune auto mode with distributed cache coordination and persistent tune/load configs

Validation

• pre-commit run --all-files
• Python bytecode compilation for lightx2v
• JSON parsing and shell syntax checks for the added configs and launchers
• 2-GPU FlashInfer TP smoke test covering tune -> save -> cold load
• full 8-GPU TI2I TP + FlashInfer runs: first run persisted 18 autotune configs; a second cold run loaded those configs in both AR and diffusion and completed all 50 denoise steps

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.