For example even with live wcfs lsof +D output differes from regular lsof and is misleading:

    LSOF /dev/shm/wcfs/b53b61099c740b452b383db6df6dce4ad6d23ba2:

    COMMAND     PID       USER   FD   TYPE DEVICE SIZE/OFF NODE NAME
    runUnitTe 29137 slapuser34   24r   REG   0,48      111    4 /dev/shm/wcfs/b53b61099c740b452b383db6df6dce4ad6d23ba2/.wcfs/zurl
    runUnitTe 29137 slapuser34   25u   REG   0,48        0    7 /dev/shm/wcfs/b53b61099c740b452b383db6df6dce4ad6d23ba2/head/watch
    runUnitTe 29137 slapuser34   26r   REG   0,48  2097152    9 /dev/shm/wcfs/b53b61099c740b452b383db6df6dce4ad6d23ba2/head/bigfile/00000000000077e2

    LSOF +D /dev/shm/wcfs/b53b61099c740b452b383db6df6dce4ad6d23ba2:

    COMMAND     PID       USER   FD   TYPE DEVICE SIZE/OFF       NODE NAME
    runUnitTe 29137 slapuser34   24r   REG   0,48      111 4294967295 /dev/shm/wcfs/b53b61099c740b452b383db6df6dce4ad6d23ba2/.wcfs/zurl
    runUnitTe 29137 slapuser34   25u   REG   0,48        0 4294967295 /dev/shm/wcfs/b53b61099c740b452b383db6df6dce4ad6d23ba2/.wcfs/zurl
    runUnitTe 29137 slapuser34   26r   REG   0,48  2097152 4294967295 /dev/shm/wcfs/b53b61099c740b452b383db6df6dce4ad6d23ba2/.wcfs/zurl

Just aborting the FUSE connection does not make WCFS to exit. This abort
is needed to avoid deadlocks, but we should also signal WCFS it should
stop. Not doing so will lead to timeout in wait after FUSE connection
abort and `kill -QUIT` triggered, which makes WCFS crash and print lots
of traceback to stderr before exiting.

Previously that message was printed to stderr and it was very confusing:
we had something in the exception message, something in the log and
something in stderr.

Now exception message is self-sufficient and log provides more details.

This patch does to start what b0ca031f did for join/serve.

Stop duplicating the code and factor logic to "attach to wcsrv, or
prepare to start" into _try_attach_wcsrv function.

For example runUnitTest in ERP5 runs test program with $PATH set without those
directores. This leads to failures like

      File "/srv/slapgrid/slappart16/srv/testnode/djk/soft/950bed0366ff018b829e9bbb7dc3326f/parts/wendelin.core/wcfs/__init__.py", line 533, in _is_mountpoint
        mounted = (0 == subprocess.call(["mountpoint", "-q", path]))
      File "/srv/slapgrid/slappart16/srv/testnode/djk/shared/python2.7/a8bdae732e550b385eb04a06ed09d383/lib/python2.7/subprocess.py", line 172, in call
        return Popen(*popenargs, **kwargs).wait()
      File "/srv/slapgrid/slappart16/srv/testnode/djk/shared/python2.7/a8bdae732e550b385eb04a06ed09d383/lib/python2.7/subprocess.py", line 394, in __init__
        errread, errwrite)
      File "/srv/slapgrid/slappart16/srv/testnode/djk/shared/python2.7/a8bdae732e550b385eb04a06ed09d383/lib/python2.7/subprocess.py", line 1047, in _execute_child
        raise child_exception
    OSError: [Errno 2] No such file or directory

where those system programs are not found.

Implement Server.stop using most of the functionality that was
previously living in tWCFS test part.

I forgot to update go.sum entry there and it started to fail to build.
Thanks to Rafael for the bug report.

This will, hopefully, be used to fix rebuild.

* t2:
  .
  .
  .
  .
  .
  .
  .
  X treediff ret += δtkeycov
  .
  .

Hopefully rebuild will use this to correctly recompute trackSet

As a consequence rework treediff to work on key ranges instead of
individial keys.

Adjacency coverage gets wider, but that is ok, because in practice if
keys are added to a BTree - that means that something is added to
ZBigFile, and it is very likely that just added data will be accessed
again. In other words it means that wcfs will very likely track those
data on nearby future read access, so we don't loose anything if we
start to track those keys right at the beginning.

* t2: (74 commits)
  .
  .
  .
  .
  .
  .
  X kadj must be taken into account as kadj^δZ
  .
  .
  .
  .
  .
  .
  .
  .
  .
  .
  .
  .
  .
  ...