Fix missing GC root in table.c

[yuyichao]

This is the second GC (usage) bug discovered while investigating #15271. This has a higher chance to be the real cause since I can see how some recent change can put a higher pressure on the ObjectIdDict rehashing... (OTOH, it seems to be as old as 2011/2010 and I'm surprised that no one have hit it before.....)

@tkelman backport 0.4 might need to remove the dump.c part. (the change in dump.c is only a comment now so it should be easy to backport.) Let me know if it is not straight foward when doing the backport.

@colbec If you can try to see if this fixes the issue.

[yuyichao]

Local GC stress test on subarray passes 11 times (8x release, 3x debug) where as it segfaults 6 out of 6 times with debug build before. Will wait for a few more test runs and merge this later today to see if it helps #15271.

[carnaval]

thanks a lot man :-) those bugs really are the nastiest

[timholy]

Whether it fixes the problem or not, this is awesome @yuyichao.

[colbec]

@yuyichao FYI I have been running tests (about 10 today with testall) and not seen the issue. I am currently on Version 0.5.0-dev+2975 (2016-03-02 15:55 UTC) and will keep testing at a frequency higher than I would normally to see if I can see the issue again. Clearly it puts a burden on my machine and cycles, but I have noted the interest in the issue and will do what I can to support the team effort.

[tkelman]

Thanks for keeping an eye out @colbec! When I'm running tests repeatedly I like to do something like

for i in `seq 20`; do
  make testall || break
done

[yuyichao]

I usually add a mail <email address> before the break.