TestChildWatch was introduced to test what happens when the response is too large for the default
allocated buffer size. It used to create 10k nodes which could take a very significant amount of
time. The same effect can be achieved with far fewer nodes, significantly speeding up the test's
runtime. The test now runs in 5s on my machine instead of sometimes multiple minutes...

TestSetWatchers tested something similar, except that it checks that the outgoing setWatchers packet
is broken up into multiple packets when it's too large. Using a similar trick we can generate
names of specific lengths to test that the behavior is correct. It was also flaky because if your
local ZK deployment is a little slow, deleting all the nodes can take longer than the session
timeout, spuriously failing the test. This has also been fixed, and the test now runs in a little
over 5 seconds as well, instead of failing.

Finally, standardize the ZK server version checking to be a bit more flexible and friendlier towards
future versions of ZooKeeper (note: the original implementation doesn't even work because the env
variable name is incorrect... It `ZK_VERSION`, not `zk_version`)

Implements the new persistent watch types introduced in 3.6, along with some utilities that are
critical when implementing local caches.