Add instrumentation and make system robust for actor restarts

[dongwang218]

Why ?

collect more instrumentation metrics. Make system robust for actor restarts.

How ?

• instrumentation: add dequeue latency
• when ENABLE_INSTRUMENTATION is changed to true, add the instrumentation metrics as part of the orchestrator. This is to collect detailed metrics as prometheus scrapping internval is about 1 second, not detailed enough. With this, we can call collect_instrumentation_metrics.py to calculate detailed latency, throughput and networking metrics.
• Added dummy coral agents to study system bottleneck when agent is fast.
• Use a background task to refresh the list of active actors, when agent actor restarts, we proactively refresh the Ray actors to avoid stale actors. Add local caching as the up to date mapping is kept in Sink, calling Sink to obtain the list is not scalable.
• add dead_orchestrator_tracking to proactively timeout long running task in case they are dead. Use newly arrived task to timeout old ones and for late arrival, chang the timeout window to make it harder to timeout.

Test plan

Tested all commands same as #115

[yangli5t]

nit: same as the one in collect_agent_metrics.py?