Add support for multi-field keys to terms aggs

[imotov]

There is a need in kibana (see elastic/kibana#77632) to dynamically form a key from multiple ordered fields, similar to composite aggs. The example that kibana team provided is to find top 10 cpu users, where each user is identified by a compound key consisting of datacenter id, host id, and container id. The required result is 3 keys and cpu usage metric with everything sorted by cpu usage descending:

datacenter	host name	container id	cpu (sorted 🔽)
us-east-1	f80a10	b91e34	55.20
us-east-2	6dd558	493fb5	42.00
us-west-2	ba4d4	9675e5	40.73
us-west-1	60413e	0574ef	39.92
us-east-1	f80a10	b91e34	36.21
us-east-2	6dd558	47b9c81	34.58
us-west-1	ba4d4	9675e5	24.01
us-east-1	f80a10	b91e34	22.40
us-east-2	6dd558	493fb5	15.91
us-west-1	27380b	f44031	12.33

The current solution is really awkward and limiting concatenation of fields using a script. We would like to offer a more streamlined solution for Kibana here.

[elasticmachine]

Pinging @elastic/es-analytics-geo (Team:Analytics)

[dgieselaar]

I think this will be very useful for APM as well. There's a lot of data that usually should only be aggregated together based on a set of fields. E.g., for transaction duration data, we want to aggregate on service.name, service.environment and transaction.type. Or service.name and service.node.name for JVM memory usage.