[Store] Throttle repeated no-master client logs via LOG_EVERY_T (follow-up to #2484)

[silas-scitix]

What

This is a follow-up to kvcache-ai#2484. During a persistent no-master window, the
storage heartbeat loop (StorageHeartbeatThreadMain) and the leader
monitor loop (LeaderMonitorThreadMain) re-emit the same failure log on
every retry (fail_ping_interval_ms / kErrorRetryInterval, ~1s). With a
large number of clients this becomes a log storm that drowns out other
diagnostics and grows the log volume linearly with the client count.

This wraps each repeating no-master log site with glog's built-in
LOG_EVERY_T(SEV, 30), so each call site emits at most once per 30
wall-seconds per process while keeping the original message text,
severity, and the first occurrence. No throttle helper is hand-rolled;
the built-in glog macro is used directly.

Sites changed (mooncake-store/src/client_service.cpp)

StorageHeartbeatThreadMain (heartbeat ping/reconnect loop):

• "Failed to ping master"
• "Failed to get new master view"
• "No active master view is published yet"
• "Failed to connect to master"
• "Failed to ping master for N times (non-HA)"
• "Reconnect failed to"

LeaderMonitorThreadMain:

• "Failed to wait for leader view change"
• "Failed to switch to leader"

Not changed

The one-shot connect logs in ConnectToMaster are intentionally left
unthrottled (they fire once per connect attempt, not in a retry loop).

Behavior

Behavior-neutral except for log frequency. Retry intervals, sleeps,
ping_fail_count, and reconnect control flow are untouched; only the
LOG() call at each repeating site is swapped for the throttled glog
macro.

Testing

• Built GREEN on a fresh upstream main checkout (7e9bdde).
• scripts/code_format.sh --check passes (clang-format clean on the
  changed file).
• ctest: no regression on the directly relevant tests
  (non_ha_reconnect_test, ha_recovery_test,
  ha_backend_availability_test, client_integration_test,
  master_service_test, client_metrics_test).