…er events

Replace LastEvent tracking with marker-based queue synchronization:
- Remove LastEvent_ member and updateLastEvent() from Queue class
- Add addDependenciesQueueSync() template that creates marker events in
  other queues for implicit synchronization between blocking/default streams
- Make query() pure virtual, backends implement using native sync APIs

Level Zero changes:
- Use ZE_COMMAND_QUEUE_FLAG_IN_ORDER for command queues
- Implement query() with zeCommandListHostSynchronize zero-timeout check
- Move event maintenance from EventMonitor to on-demand in getEventFromPool()
- Add createEventDedicated() for callback events (workaround for PVC driver
  issue with event reuse between immediate and regular command lists)
- Get dependencies before locking CommandListMtx to avoid deadlock
- Reduce event pool size from 1000 to 10

OpenCL changes:
- Track IsEmptyQueue_ to handle query() for never-used queues
- Implement addDependenciesQueueSync() with cl_event markers
- Rework callback mechanism to use marker enqueued from callback thread
  instead of user events (workaround for Intel driver bug)

Other:
- Add checkEvents() virtual method to Context for event processing
- Fix Device destructor to avoid holding multiple locks simultaneously
- Remove L0CollectEventsTimeout env var (no longer needed)
- TestBufferDevAddr now conditional on OpenCL backend

Callbacks now use shared event pool via createEventShared, so the
dedicated event creation function is no longer needed.

…nization

This change flushes the other queue before using the marker event as a dependency, addressing synchronization issues on certain OpenCL drivers (e.g., Arm Mali). This ensures that the marker event is properly submitted and can be utilized across different queues.

This update introduces a call to clFlush in the addCallback method to guarantee that callbacks are triggered correctly on drivers that require explicit flushing, such as ARM Mali. This change enhances the reliability of event callback execution in the OpenCL backend.

Restore createEventDedicated() and use dedicated events for callbacks
to avoid Intel driver bug. Track GpuReady event so event monitor can
properly check completion status. Remove unused dedicated callback
command list code.