Refactor unified trainer async flow and add precomputed-advantage + Tinker rollout fixes

[listar2000]

Summary

This PR streamlines the experimental trainer execution path, standardizes episode/task metadata handling, and adds support for workflows that provide pre-computed per-
token advantages. It also includes several Tinker integration fixes (rollout parsing, model length wiring, and resume behavior).

What Changed

Unified trainer lifecycle

• Removed the background event-loop thread / thread-safe coroutine wrapper flow.
• Kept fit() as sync entry point, now using asyncio.run(...).
• Added explicit async entrypoint (fit_async) and moved workflow-pool initialization into async startup.
• Passed use_precomputed_advantage through trainer setup.
• Switched validation UID grouping to use normalized episode.task_id.
• Updated tracking payload to serialize self.config (full config object used by trainer).

Agent / episode model cleanup

• Added Step.from_model_output(...) to centralize Step construction.
• Updated completers to use the shared constructor.
• Relaxed Step.prompt_ids typing to allow non-int prompt blocks used by some backends.
• Added cached Episode.task_id and Episode.rollout_idx.
• Switched trajectory-group identifiers/properties to cached-property usage.

Advantage collection behavior

• Added algorithm.use_precomputed_advantage (default false) to config/dataclass.
• When enabled, collector consumes per-token step.advantage lists directly.
• Missing/mismatched per-step advantages are defaulted to zeros with warnings.
• Scalar step.advantage in precomputed mode now raises an error.
• In normal RL mode, advantages are still computed from rewards; logs warning if precomputed values are present and being overwritten.

Transform / metadata consistency

• Replaced ad-hoc episode.id.split(":") usage with episode.task_id / episode.rollout_idx.
• Updated metadata emitted for visualization/transform paths accordingly.

Tinker backend and rollout fixes

• Wired training.max_length into Tinker engine as max_model_length.
• Added optional rollout_engine.renderer_name config.
• Corrected stop-sequence fallback to use tokenizer eos_token_id when available.
• Improved sampling-param handling (max_tokens precedence and copy semantics).
• Added backend-specific completion-id extraction in TITO completer.
• Added parser-bypass response parsing path when configured.
• Added async compute_logprobs(...) API to Tinker engine.
• Added resume support from training.resume_from_tinker_id (tinker://.../weights/<batch>).