WCFS allows issuing simultaneous watch requests and when two watch
requests are simultaneously issued for the same file there was a race in
their handling: the code was relying on w.atMu.W to protect setupWatch
from concurrent readPinWatcher, and also, seemingly from another
setupWatch running on the same file.

But there is a bug about that: lacking atomic primitive to downgrade
RWMutex from wlock to rlock, atMu.W was first fully unlocked and then
rlocked again. The code prepare wrt readPinWatcher to start running in
that unlock->rlock time window, but it was not prepared wrt another
setupWatch starting to run on the same file in that pause time.

-> Fix that via using dedicated Watch.setupMu lock that protects
   setupWatch from setupWatch.

Test is, hopefully, TODO.

My mistake from 6f0cdaff (wcfs: Provide isolation to clients)

/reviewed-by @levin.zimmermann
/reviewed-on !18