Optimize insecure registry failover by caching failover history

[chanseokoh]

Fixes #946.

Significantly cuts the number of HTTP registry connections, by more than half in my testing from 528 to 248. Once a failover occurs for a registry host, it is not repeated (down to 4 from 127 as shown below).

In practice, however, due to Jib's concurrent connections, it may repeat the failover for the same host but mostly only once or twice.

$ grep 'User-Agent: jib' network_traces.sh.log | wc -l
528
$ grep '\(again with no TLS verification\.\)\|\(again with HTTP\.\)' network_traces.sh.log | wc -l
127

$ grep 'User-Agent: jib' network_traces.sh.log | wc -l
248
$ grep '\(again with no TLS verification\.\)\|\(again with HTTP\.\)' network_traces.sh.log | wc -l
4

Increases the failover log level from INFO to WARNING.

[chanseokoh]

This is now ready for review.

[chanseokoh]

Huh, this branch isn't really the one that I originally intended to put up. Somehow I see some commit is lost. The failoverHistory map should be a ConcurrentHashMap.

[TadCordle]

Oh and changelog, haha

[loosebazooka]

We can probably put this out with 1.8...