Experimental change to not care about cversion for Watched collections with persistent recursive watch

[patsonluk]

Description

An experimental branch/PR to explore using persistent recursive watch without maintain the versioning of PRS entries (whether it's cversion or pzxid as proposed in #105 )

There are various challenges in #105, and one of the biggest issues was to keep the versioning of PRS entries. So this PR explores whether it's possible to ignore such versioning for "watched" collections. As watched collections are supposed be updated by watched events/notification (and that ONLY after the initial fetch), therefore versioning is not important to watched collections

This branch passed both test cases in PerReplicaStatesIntegrationTest and TestPerReplicaStates. Ran PerReplicaStatesIntegrationTest#testZkNodeVersions (exclude the non PRS collection) and captured metrics similar to https://fullstory.atlassian.net/browse/SAI-4544?focusedCommentId=259534.

It's found that this PR improves (ie reduce fetches) for collection creation, add/remove replica. but for node restart, it's pretty much the same as 9.2. This is probably due to the fact that most fetches for restart are actually triggered from ZkStateReader#waitForState (

fullstory-solr/solr/solrj-zookeeper/src/java/org/apache/solr/common/cloud/ZkStateReader.java

Lines 1955 to 1965 in 74984f5

	registerDocCollectionWatcher(collection, watcher);
	// wait for the watcher predicate to return true, or time out
	if (!latch.await(wait, unit))
	throw new TimeoutException(
	"Timeout waiting to see state for collection="
	+ collection
	+ " :"
	+ docCollection.get());
	return docCollection.get();
	} finally {
	removeDocCollectionWatcher(collection, watcher);

) which for each core, register/unregister Doc collection listener multiple times.