Fix monitoring: count retry attempts correctly and use K/M/B token units

[liuw1535]

Motivation

• Monitoring reported each retry as a successful request because retry failures were not tracked separately from the final response, causing totals to be inflated; token unit formatting in the dashboard used Chinese large-number units instead of English compact units.

Description

• Change src/utils/usageStats.js to accept successCount/failedCount, add toSafeCount(), and update record() to aggregate per-day and per-model requests, success, and failed using those counts.
• Add recordUsageAttemptFailure(req, ...) helper and wire it into the retry flow so failed retry attempts are accumulated on req.apiUsageMetrics instead of being mis-counted as successes.
• Call the new onAttemptFailure callback in src/server/handlers/common/retry.js so callers can record failed attempts, and add onAttemptFailure to createRetryOptions in src/server/handlers/openai.js, src/server/handlers/claude.js, src/server/handlers/gemini.js and src/server/handlers/geminicli.js to invoke recordUsageAttemptFailure(req, { model }) on retry failures.
• Update the request-tracking middleware in src/server/index.js to pass successCount and failedCount into usageStats.record() based on req.apiUsageMetrics.failedAttempts and the final res.statusCode so totals reflect conversation attempts (e.g. 3 failures + 1 success → total 4).
• Change frontend token formatting in public/js/monitor.js to use formatTokenNumber() with K/M/B abbreviations and apply it to Tokens card, model legend and donut center.

Testing

• Ran static syntax checks: node --check on src/utils/usageStats.js, src/server/index.js, src/server/handlers/common/retry.js, src/server/handlers/openai.js, src/server/handlers/claude.js, src/server/handlers/gemini.js, src/server/handlers/geminicli.js and public/js/monitor.js, which completed without errors.
• Attempted npm run build:linux to validate packaging, but pkg failed to fetch Node base binaries/source (remote 403 Forbidden), which is an environment/remote-cache issue unrelated to the code changes.

---

Codex Task