Add Dream-7B model support

[iOptimizeThings]

Background

Dream-7B is a 7B-parameter diffusion language model from the Dream team (HKUST + Huawei). Architecturally similar to LLaDA - masked diffusion, bidirectional attention, parallel commit - but uses a different commit policy: entropy-based rather than confidence top-k.

Investigation during v0.1 development found that the default top-k confidence commit causes Dream to emit empty strings (content: "") when max_tokens ≥ 256. The model's confidence distribution doesn't match LLaDA's: high-entropy positions dominate and the top-k selection picks EOS cascades. At max_tokens ≤ 128, temperature=0, Dream works correctly.

Acceptance criteria

• Add dlmserve/models/dream.py loader (analogous to llada.py)
• Add entropy-based commit policy: commit_top_k_by_negative_entropy() in sampler.py
• Engine dispatches commit policy by model family (dlmserve/engine.py)
• All 4 LocalLeap quality gates pass for Dream (or document why a gate is skipped)
• Quality re-verification (MMLU, BLEU, determinism) - per CLAUDE.md rule about quality after acceleration
• README "Supported models" table updates to ✓ v0.1.1
• examples/chat.py --model Dream-org/Dream-v0-Instruct-7B produces coherent output at max_tokens=256

References

• Dream model card: https://huggingface.co/Dream-org/Dream-v0-Instruct-7B
• Dream paper: https://arxiv.org/abs/2510.15968
• v0.1 investigation notes: see docs/adrs/ for the EOS-cascade analysis