nsqd: race when persisting metadata

[mreiferson]

since Notify() can be called from many goroutines this needs to Lock not RLock cc @jehiah

[jehiah]

i think this change is good; but it makes me think that another side of this is that we shouldn't use a constant name for the "%s.tmp" file. That would allow our generation to work, and the atomic rename to always succeed regardless of how many time this happens or doesn't happen in parallel.

[mreiferson]

discussed offline... because syscalls are scheduler yield points, we cannot rely on just the atomic rename to provide global ordering of concurrent metadata writes (aka, we need this lock)