refactor: make orchestration verifiers v1-native

[hallerite]

Stacked on #3219; merge that PR first.

Targets prime-rl 0.8.0 and the checked-out verifiers v0.3.0 tag.

Summary

• make the typed v1 Episode the sole orchestration currency from dispatch through sinks, batches, metrics, and persistence; remove EpisodeResult, copied episode_ok state, and parallel episode/rollout buffers
• derive flattened trace, effective-subset, and per-environment views from their parent episodes while preserving zero-trace failures and multi-agent episodes
• make scheduling and batching episode-native: derive capacity from in-flight work, allow groups larger than capacity, preserve complete task groups at trainer-batch boundaries, and separate episode failures from trace errors
• tighten v1-only lifecycle and configuration around canonical taskset iteration/head(), num_tasks, train-only source ratios, client shutdown, and the v1 renderer client
• update configs, docs, metrics guidance, and the monitoring skill for the episode JSONL schema and per-agent eval semantics

Validation

• ruff check passed on the changed Python surface
• ruff format --check passed on the changed Python surface
• focused orchestrator unit suite — 72 passed
• focused config validation — 2 passed
• dispatcher/train-sink/eval-sink import smoke passed
• successful v1 episode conversion and record round-trip smoke passed

The optional Qwen3-VL/vLLM integration test was migrated to the v1 TrainClient path and linted, but was not executed in this local macOS environment.