Allow FUNSD benchmark to customize budget allocation

[warrofua]

Summary

• add allocator_kwargs to the FUNSD benchmark helper and thread it through the shared evaluation path
• instantiate a RegistryAwareBudgetAllocator once per evaluation and reuse it for primary and ablation bottlenecks
• extend RegistryAwareBudgetAllocator to accept optional field weight and minimum quota overrides
• fix ablation bottleneck instantiation so each ablation receives a shared allocator-backed bottleneck
• log compression records for benchmark usage events so type checking can resolve CompressionRecord

Testing

• pytest tests/bottleneck -q
• mypy .

---

https://chatgpt.com/codex/tasks/task_e_68fce757359c832c9943dce3b04f4118