replica set not configured when read preference is secondary

[linearb]

When making a connection with read preference secondary, it does not correctly track
the replicaset members in __allnodes. The problem seems to be that the task looping call
to _Connection.configure() only runs once because it throws an AutoReconnect exception
on line 136 the first time it is called. However, at least in the environment I tested, the first
(and only) run of the task looping call happens before the connection is made.

[IlyaSkriblovsky]

This seems like a legacy behavior from ancient times when ismaster was only used to check if node is primary or secondary. Later configure was modified to extract all node addresses from ismaster response, but condition under that configure is called was not modified. This needs to be fixed.