Fix Deep Review interruption recovery and reviewer continuation behavior

[limityan]

Summary

This PR improves Deep Review recovery after failures, policy-ineligible reviewers, and user-initiated stops. It keeps recovery actions in one coherent action row, supports continuing manually stopped Deep Reviews, and avoids restarting optional or non-applicable reviewers during continuation.

Changes

• Simplify the Deep Review recovery action bar:
  • Remove the standalone recovery-plan toggle.
  • Keep Continue review, model settings, diagnostics, and partial-results actions in the same action area.
  • Minimize the recovery panel after a continuation request starts successfully.
• Support continuation after a user manually stops a Deep Review.
• Preserve completed reviewer work during continuation.
• Mark policy-ineligible or non-applicable reviewers as skipped so they are not rerun.
• Apply conditional ReviewFrontend scheduling for explicit non-frontend targets.
• Mark built-in ReviewFrontend as a review agent so valid frontend review tasks pass policy validation.
• Ensure subagent sessions publish their active turn state so stop actions can cancel the correct turn.
• Add Deep Review prompt guidance to keep locale/i18n-only reviews scoped and avoid repeated tiny reads.

Testing

• pnpm run type-check:web
• pnpm run lint:web
• pnpm --dir src/web-ui run test:run
• cargo check -p bitfun-core
• cargo test -p bitfun-core deep_review -- --nocapture
• git diff --check