feat: adapt orchestrator to index-addressed env server

[mikasenghaas]

Summary

Companion to verifiers PR #1829, which makes the v1 env-server index-addressed with no required task count so it can serve lazy/unbounded tasksets — InfoResponse.num_tasks is now int | None (None = served lazily; the caller drives task_idx). This adapts the orchestrator to that API.

• Env.num_tasks is now int | None (None when the taskset is served lazily).
• EvalEnv.start requires num_examples to bound a lazily-served eval — an unknown/infinite count can't be enumerated (raises a clear error if num_examples < 0).
• TrainSource:
  • Finite path unchanged — enumerate the env's index range, shuffle, reshuffle each epoch, weight by dataset size.
  • Unbounded path (num_tasks is None) — next_example streams a monotonically increasing task_idx (no epoch/reshuffle; the taskset's own generator, e.g. RNG-seeded pairs, supplies the variety).
  • An unbounded env has no dataset size to weight by, so it requires an explicit ratio (clear error otherwise).

Backward-compatible: against a server that still reports a count (every current taskset that returns a list, and the v0 bridge), every env is bounded and behavior is identical. The change is safe to land before the verifiers pin bump.

Verification

Unit-checked TrainSource against stub envs:

• finite env → epochs cover all indices, reshuffled (e.g. [2,0,1,3, 1,2,0,3, ...]);
• unbounded env (num_tasks=None, with a ratio) → streams [0,1,2,3,4,5];
• unbounded without a ratio alongside a finite env → ValueError (no size to weight by);
• mixed finite + unbounded with ratios → both envs sampled.

ruff check / ruff format clean (pre-commit).

Follow-up

The verifiers pin bump (to the merged #1829 commit) lands separately once that PR merges; only then does an actual unbounded taskset report num_tasks=None end-to-end.

[mikasenghaas]

Superseded by #2882, which reimplements the same orchestrator change off main (this branch was based on feat/nano-as-v1).