close SummaryWriters

[pkol]

SummaryWrites should be closed after use. Otherwise they have some
threads running and prevents application from terminate. Relaying on
garbage collector for this is unstable - sometimes it kills them faster,
sometimes slower.

This fixes #298

Without this commit, following command does not end (process does not exit):
python rl_trainer.py --config_file=rl/configs/light_awr_cartpole.gin --output_dir=/tmp/awr1

[afrozenator]

Hey @pkol this is really helpful! Will merge it in shortly, may I ask how you debugged this issue? i.e. come to the conclusion that the SummaryWriters were the issue

[pkol]

Hi @afrozenator .
Thanks for accepting my contributions :-)

I was debugging it for a while... Finally I printed traceback of all threads which remains after RL training (using faulthandler.dump_traceback() ) and there were only threads running code from event_file_writer.py (so I suspected SummaryWriters). Then noticed there is also SummaryWrite.__del__() calling close() from EventFileWriter. So I checked if calling 1gc.collect()1 after training will help. And it did.
I think it was something like above but it's really hard to recall this process exactly.

BTW there is still some mystery. In TF 1.5 in _EventLoggerThread.__init__() there is self.daemon=True so this thread should not block app from exiting. However it's set after calling Thread.__init()__ which I think (but I'm not 100% sure) is a bug.

[afrozenator]

Thanks @pkol for teaching me something new today! Merging in shortly.