Add LLaDA 2.0 (inclusionAI) model support

[iOptimizeThings]

Background

LLaDA 2.0 (by inclusionAI, Ant Group) is the next generation of LLaDA, published after the gsai-ml LLaDA-1.5 release supported in v0.1. Despite the shared "LLaDA" name, LLaDA 2.0 uses a block-diffusion architecture (related to BD3-LMs) rather than the pure masked-diffusion contract of LLaDA-8B/1.5.

This is not a drop-in addition. Block diffusion changes:

• Forward pass shape (block-wise attention masks)
• The denoising loop (per-block iteration is structural, not a config flag like block_length)
• Quality gates (different reference implementation to bit-match against)

Acceptance criteria

• Vendor the inclusionAI reference loop into reference/
• Add dlmserve/models/llada2.py loader
• Implement block-diffusion denoise path in dlmserve/denoise_loop.py (or new file)
• Reference-match gate at temperature=0 (token-exact vs the vendored reference)
• All 4 LocalLeap quality gates pass (or document scope reduction)
• LocalLeap calibration run for LLaDA 2.0 (scripts/calibrate_local_leap.py --model inclusionAI/LLaDA-2.0-...)
• README + perf_log updates

References

• inclusionAI org on HuggingFace: https://huggingface.co/inclusionAI
• LMSYS day-0 LLaDA 2.0 post: https://www.lmsys.org/blog/2025-12-19-diffusion-llm/
• BD3-LMs paper (related architecture): https://arxiv.org/abs/2503.09573