Replace mongos high availability with load balancing

Aspect of the Server Selection Spec implementation: if given a seed list of mongos addresses, MongoClient must use evenly the nearest one and all those <= latencyThresholdMS farther.

Replaces prior behavior called "mongos High Availability" or "HA", where MongoClient compared all mongos seeds' latencies and pinned to the nearest one until a network error, then checked their latencies anew.

This was done or nearly so when we deleted requests and pinning. Verify the work was completed, re-enable and fix mongos HA tests (including in test_ha.py) so they test load-balancing instead, verify that latencyThresholdMS is respected.

Update docs.