1.3.2: Optimize stale subscription cleanup rake task

Fixed

• Cleanup tasks are much faster and use bounded memory: SCAN/SSCAN/ZSCAN are called with a configurable COUNT (new redis_scan_count setting, default 1000) instead of Redis' default of 10, members are read incrementally with SSCAN/ZSCAN instead of SMEMBERS/ZRANGE (which materialized entire sets in the rake process), and the per-member EXISTS, OBJECT IDLETIME, SREM, ZREM and DEL calls are pipelined in batches instead of being issued one round trip at a time. [@jjb] (#56)