Support stable iteration detection

Stable version in many aspects

Provide much more accurate timestamp of each perf-event

1. great improvement on timing alignment of gpu/cpu/net traces.

2. cuda api tracing is added.
3. metrics like instructions and cache-misses of events are added.

periodical test passed

support correct CPU/GPU time alignment and iteration detection

many updates and only use python3.6 from now on

Now only tcpdump and profile-all-cpus need privilege of root.

Support better visualization info.

Better timestamp alignment method and optional CPU coverage