zk-ruby/zookeeper

Previous URL 404s.

This is necessary if the default linker search path does not include
the shared libraries for the version of Ruby being used.

Previous approach was to use polling with a period of 100 msec. The
new approach is to associate a pipe with the queue that we use to
send events from ZK to Ruby land. When the Ruby consumer runs out
of events to process, it does rb_thread_select() on the pipe.
Whenever the ZK side enqueues a new event, we write into the
pipe.

* Previous coding depending on recursive mutexes; since we can't
  easily guarantee those and use the static PTHREAD_MUTEX_INITIALIZER,
  just avoid the need for recursive mutexes for now.

* Add a missing pthread_mutex_unlock() to an error return path.

* Fix harmless compiler warning.

Conflicts:
	ext/zookeeper_lib.c

The current coding pre-allocates the queue node for the next
to-be-inserted queue element. This seems a strange way to
do things, but I'll leave things as they are for now.

The check ZKH_SETUP_EVENT was incorrect for the zkrb_state_callback. req_id
can be different from the global request id and still be using the same
callback function. Nasty work. Fixes #1.

ZK allows a path to be created that has no associated data. In
the previous coding, doing a #get() on such a path would result
in raising a Ruby exception: zoo_get() returns data_len == -1
for the path, which isn't a legal string length.

This patch fixes a per-connection memory leak: the previous
coding allocated a zkrb_calling_context for each ZK handle
but neglected to free it.

This patch also raises a Ruby exception if any operations
on performed on a closed ZK handle, until #reopen() is
called. Such operations would just segfault anyway.

If #close() was invoked, zookeeper_close() might be invoked
again by the Ruby GC, leading to a double free(). The previous
coding also neglected to call zookeeper_close() on a
non-connected ZK handle, which would leak resources.

Previously assigned to the wrong union member -- not super likely
to cause a problem, but illegal per C99 strict aliasing.

strlen() + strcpy() is silly.

Not clear that this is the right fix, but unbreaks the build
for me.

Cleans up one possible SIGSEGV.