[Refactor] Centralize Fast inference attempt diagnostics

[roomote-roomote]

​Created by Roomote. View the task or mention @roomote-roomote for follow-up asks.

What changed

• Added one diagnostics-owned inference attempt API that records context submissions, provider retries, model and prompt milestones, outcomes, timing, and bounded failure logs.
• Reduced the Fast turn service to execution and retry-policy callbacks instead of maintaining duplicate attempt counters, stage derivation, timing, and telemetry fan-out.
• Added focused lifecycle coverage for success, provider recovery, setup and generation failures, clean retries, and side-effect recovery while retaining the existing telemetry schemas.

Why this change was made

Fast turn execution and turn diagnostics independently tracked the same inference lifecycle. Keeping those parallel state machines synchronized made context and retry observability harder to extend safely.

Impact

There is no intended user-facing behavior change. Fast inference diagnostics now have one owner, preserving privacy-safe telemetry and retry semantics while reducing orchestration inside the main turn service.

[roomote-community]

No code issues found. See task

_{Reviewed f38c710}