I believe this was exposed by 5492dc20, because the sending of the 500
would have already failed and prevented this logging. On Python 3,
this will be a ConnectionResetError, which is a subtype of OSError,
which is why we check for both OSError and socket.error.
Bonus: this fixes a race in test-hgweb.t where sometimes the
ECONNRESET wouldn't happen, because now we just don't log those
errors.