MGM: Traffic Shaping - Improve locking

Fix deadlock in Start()/Stop(): both are called from Enable()/Disable() via
FsView::ApplyGlobalConfig() while IConfigEngine::ApplyConfig() holds the
ViewMutex write lock. SyncTrafficShapingEnabledWithFst() acquires a read lock
on the same thread causing a deadlock. Removed it from Start()/Stop(); the
call in TrafficShapingEngine::ApplyConfig() (after the lock is released)
provides the immediate FST notification on startup.

Other fixes:
- Minimize ViewMutex hold time in UpdateLimits() and
  SyncTrafficShapingEnabledWithFst(): snapshot online node names under a brief
  read lock, then re-acquire per-node for SetConfigMember
- Release mMutex before SetGlobalConfig/AutoSave in all Set/RemoveXxxPolicy
  methods to avoid holding it during slow config-engine I/O
- Snapshot mAppPolicies under a shared lock in UpdateTrafficShapingController()
  to fix a data race with concurrent Set/RemoveAppPolicy calls
- Move mMutex acquisition before JSON parsing in LoadPoliciesFromString()
- Replace sleep_until with assistant.wait_for() in all three background threads
  so termination requests wake them immediately
- Join mFstTrafficShapingConfigUpdateThread in Stop() — it was started but
  never joined, leaving it running after stop
- Add Clear() call to TrafficShapingManager destructor
- Fix UB cast in CalculateDelayUs: clamp to zero before casting to uint64_t
- Drop oldest reports instead of clearing the queue on overflow

Author: lobis