Fix unseeded generation when seed_base is unset; make the critic's request params configurable
Three fixes to agentic_upsampling/, one per commit so they can be taken separately.
Seed base. With seed_base=None the generator got seed=None, so every iteration re-rolled the noise. Best-of-N then compared images that differed by prompt and noise together, so a score change couldn't be attributed to the rewrite and runs weren't reproducible. Now one base is drawn per item and reused across iterations (slot j gets base + j), matching how an explicit seed_base already behaves. The value is recorded on the result so a run can be reproduced. A resumed run draws a new base; happy to persist it instead if you'd rather.
Critic request params. VLMQualityJudge built its request with fixed parameters, so callers reproducing a specific setup couldn't pin reasoning_effort. Adds an extra_body passthrough, default None, so requests are unchanged unless set. Also raises the critic's max_tokens to 16384: thinking tokens come from the same budget, and we measured up to ~6.7k for gemini-3.1-pro at reasoning_effort=high.
Flaky test. Samples in an iteration are generated concurrently, but the seed assertion compared the recorded list in order. It fails about 1 run in 12 on main today. Now compares each iteration's seeds as a sorted group.
17 tests pass, 15 consecutive full-suite runs clean.