fix(credential_pool): sync provider endpoints from auth state

[DanielLSM]

Problem

When a provider has multiple API surfaces (e.g. Z.AI regular vs coding-plan endpoints), the pool entry's base_url can become stale after runtime endpoint detection finds a different working URL. The entry stays exhausted on the wrong endpoint indefinitely, even though auth.json contains the correct detected endpoint.

Changes

agent/credential_pool.py

• Add _sync_endpoint_from_provider_state(): generalized method that checks provider_state.<provider>.detected_endpoint in auth.json and adopts a newer endpoint when the cached base_url differs from the pool entry. Clears exhaustion immediately so the entry can be retried.
• Re-probe on cooldown expiry: after clear_expired resets an exhausted entry, call the provider's endpoint detection to verify the cached base_url is still valid. If the endpoint changed, update it. If no endpoint works, keep the entry exhausted and preserve the original error context.
• The sync pattern is provider-agnostic (currently enabled for zai; extensible to any provider that caches detected_endpoint in provider_state).

Related Issues

• Fixes credential pool exhaustion causing wrong endpoint fallback (pattern similar to Transient 401 blacklists credential for 24 hours — breaks single-key setups #5668 transient-auth cooldown issues and Credential pool key mix-up when multiple custom providers share the same base_url #19083 custom-provider key mix-ups).

Test Plan

python -m pytest tests/agent/test_credential_pool.py -q

All tests pass.

[DanielLSM]

Closing to reopen with clean commit history.