feat(admin): LLM token-usage analytics aggregation endpoints

[Ar1anit]

Summary

Follow-up to #522 (anonymous per-feature token-usage tracking). Once LlmUsageEvent rows exist, build read-only admin analytics aggregation endpoints so we can see how users interact with AI features in terms of tokens and cost.

Depends on #522 being merged (the data source).

Motivation

• Surface the data captured in feat(llm): anonymous per-feature token-usage tracking #522 without exposing raw rows.
• Answer questions like: tokens/cost per feature per day, per tier, per language; busiest features; cost trend over time.

Scope

• Aggregation queries over LlmUsageEvent (group by feature, tier, language, day) — sums of promptTokens/completionTokens/totalTokens/estimatedCostUsd, plus call counts and success rate.
• New admin-only endpoints under /api/v1/admin/llm-usage/*, gated by the existing ADMIN_EMAILS allow-list (same pattern as the other /admin/* routes).
• Time-range + group-by query params.
• No per-user drill-down that could de-anonymize (only aggregates, or grouped by actorHash at most for cohort sequences — never resolvable back to a User).
• Docs sync (ARCHITECTURE.md, README.md, .github/copilot-instructions.md).

Out of scope

• Frontend dashboard UI (can be a further follow-up).
• ML model (tracked separately).

Acceptance criteria

• Aggregation endpoints return correct sums/counts over LlmUsageEvent.
• Endpoints are admin-gated and return 403 for non-admins.
• No endpoint can map an actorHash back to a real user.