improvement: faster bootstrapping with start-latest=true with full picture of known consumers

when burrow starts it can read offsets from the beginning or the end:

https://github.com/linkedin/Burrow/wiki/Consumer-Kafka

```
start-latest | boolean | no | false |
If true, start consuming the offsets topic from the last (latest) offset.
Otherwise, start from the earliest offset.
```

The advantage of start-latest=true is that burrow will report lag/statuses of consumers that are actively committing fairly quickly. But if a consumer is not committing offsets, it may not be known to burrow and a status may not be available. A burrow restart can lose the "now" view of an inactive consumer.

The advantage of start-latest=false is it will read the entire __consumer_offset topic and build a full picture of all consumers the cluster is storing information on. But that can take a long time depending on how many consumer groups you have, the frequency of commits, and the configuration of the compacted topic. For us this can range from a few minutes to over half an hour during which time burrow reported status may be inaccurate.

but a client can ask the broker for consumer groups and current offsets very quickly, because the brokers cache the latest info: I wonder if this can be improved:

https://godoc.org/github.com/Shopify/sarama#Broker.ListGroups
https://godoc.org/github.com/Shopify/sarama#Broker.DescribeGroups

potentially a new burrow bootstrap could be
* start consuming __consumer_offsets from latest. but don't process anything yet.
* fetch list of groups.
* fetch the details of each group that matches blacklist/whitelist to build the now picture
* then start the processing to keep the picture updated.

Providing both the speed of start-latest=true, and a picture of all consumer groups for "now".

Start by tracing Burrow's start-latest bootstrap behavior and reading the Sarama Broker.ListGroups and Broker.DescribeGroups APIs linked in the issue. Compare their group and offset information with the blacklist/whitelist requirements; done means startup is fast while the initial now view includes matching known consumer groups and ongoing offset processing remains current.

Written by the indexing model from the issue text.