Commit d0fca860 introduced big lookup lock with the idea to "make sure
we don't return forgotten nodes to the kernel". However this way it also
started to prevent several Lookup handlers to be running simultaneously,
which can deadlock if e.g. Lookup handler somehow synchronizes with
other thread which caused the lookup:

https://github.com/hanwen/go-fuse/commit/d0fca860#commitcomment-37772099

On the surface the fix is easy: if we need to prevent lookups running in
parallel to forget, we can turn lookupLock into shared one wrt Lookup,
but exclusive wrt Forget.

However a more correct fix would be to review nodefs locking completely:
there is already per-filesystem treeLock RWMutex, which is _already_
used to synchronize for forgets and lookups. From this point of view
lookupLock is unneeded and the correct fix that d0fca860 should have been
doing is to correct the scope of treeLock to be used more precisely.

However that would be a more intrusive change, and given that nodefs is
considered deprecated today, imho it is better to proceed with making
lookupLock a shared one.

The test that is added deadlocks without provided fix.

I suggest not to reject this patch based on rationale that "nodefs is
deprecated" as there still are real filesystems that use nodefs.

I had not reviewed v2 go-fuse/fs locking yet.

Change-Id: I18e01457f474dea31dc17186dfe6db582c2e6337