Fix LDAP backoff retry logic for serverdown to actually use new connections

[rvazarkar]

The current implementation was re-using the cached, invalid connection repeatedly and not actually creating a new one. The new implementation takes multi-consumer into account, properly creates new connections, and cleans up some bad logic around retry implementation.

[ddlees]

The semaphore makes the connection retry logic a little hard to reason about but the comments really help and it sounds like y'all are planning on refactoring this to be more async.

I left one nit but you can take it or leave it.