Optimize client metadata handling with a large number of partitions

Credit to lbradstreet for profiling the producer with a large number of partitions.

There was a regression in 2.3.0 that made the client metadata handling significantly less efficient when dealing with a large number of partitions. There was a less significant regression in 2.2.0.

Lucas benchmarked the producer, but the code is shared with the consumer. See the PR for more details.