Document retryable errors a watch future might throw

In particular, a watch might throw `cluster_version_changed` during an upgrade, which a client should probably retry, but the client might not be waiting on the future within a standard `onError` style retry loop.

Review the existing documentation for watch futures and retry behavior, then identify where retryable errors are listed. Document errors such as cluster_version_changed and explain that callers may need to retry even when they are not waiting within a standard onError retry loop.

Written by the indexing model from the issue text.