fix: keep accounts available after transient network failures

[wangxiaochao19-lab]

Summary

• Treat statusless TLS/EOF/connection-reset failures as retryable connection-lifecycle errors.
• Retry these failures within the configured request round.
• Avoid reporting request-local tried-credential exhaustion as a persistent account-pool outage.
• Preserve existing handling for authentication, rate-limit, and quota HTTP statuses.

Problem

A transient upstream transport failure such as:

Post "https://chatgpt.com/backend-api/codex/responses": utls: TLS handshake: EOF

could be classified as connection_lifecycle but remain non-retryable. After the request tried the only eligible credential, selection reported a pool failure, causing Cockpit to mark the account unavailable until manual recovery.

Tests

• go test ./sdk/cliproxy/auth -run 'TestManager_ShouldRetryAfterError_RetriesStatuslessConnectionLifecycleFailure|TestManagerDoesNotReportTriedCredentialExhaustionAsPoolFailure' -count=10 — passes.
• The full auth-package run still hits the existing TestResetAuthStateIgnoresResultsFromOlderAttempts failure; it reproduces on the clean v1.3.56 baseline (cc8ea89) without this patch.

Refs #2228