feat(deep-review): add adaptive review runtime and orchestration

[limityan]

Summary

This PR introduces the adaptive Deep Review runtime and wires it into the web UI orchestration flow. It turns Deep Review from a mostly fixed multi-reviewer flow into a policy-driven review runtime that can size reviewer work, manage concurrency, surface capacity state, preserve partial progress, and present richer recovery controls to the user.

What Changed

Backend Runtime

• Add a dedicated deep_review runtime module for budget tracking, execution policy, concurrency policy, diagnostics, queue state, report assembly, shared context, task adaptation, and reviewer team definitions.
• Move Deep Review behavior out of the large policy/tool implementations into smaller runtime components with clearer ownership.
• Add adaptive strategy selection based on change risk, file count, reviewer role, timeout policy, and configured Deep Review limits.
• Track runtime diagnostics such as queue waits, provider capacity retries, effective concurrency, skipped reviewers, retry usage, and shared-context measurements.
• Add capacity-aware queue controls so reviewers can wait, pause, continue, skip, or fail fast instead of treating all capacity pressure as a hard failure.
• Preserve completed reviewer output for partial reports and recovery flows.

Frontend Orchestration

• Extract Deep Review launch, report generation, and action-bar logic into dedicated flow_chat/deep-review/* modules.
• Split the review-team service into structured defaults, strategy, types, and public service entry points.
• Add reviewer context display in task detail panels so users can see which reviewer role produced or owns a task.
• Improve timeout UI for subagent/reviewer task cards, including live timeout state and user controls.
• Add queue/capacity UX in the Deep Review action bar, including capacity wait state, pause/continue/cancel controls, partial results, recovery actions, and slower-next-time behavior.
• Update Deep Review locale strings across en-US, zh-CN, and zh-TW.

User Impact

• Deep Review can scale more intelligently for larger or riskier changes instead of launching a fixed reviewer shape every time.
• Provider or local concurrency pressure becomes visible and recoverable instead of appearing as an opaque interruption.
• Users can continue from partial reviewer results, retry incomplete slices, inspect partial output, or reduce future reviewer parallelism.
• Reviewer task cards and task detail panels expose more context about who is reviewing what, making long-running Deep Review sessions easier to understand.

Review Notes

• The backend change is primarily a runtime decomposition plus adaptive policy layer for Deep Review.
• The frontend change is primarily a module split and UX wiring for the new runtime states.
• The PR is intentionally broad because the backend runtime state and frontend recovery/action-bar state need to evolve together.