Reduce/optimize update metadata requests sent during leader election

This is part of the investigation into slow shutdowns in 0.8.1. While
logging contributes to bulk of the regression, this one also adds
quite a bit of overhead:

In addLeaderAndIsrRequest (called for every partition that is led by the
broker being shut down) we also add an UpdateMetadataRequest - each call to
addUpdateMetadataRequests does two traversals over all (global)
partitions. I think it should be straightforward to optimize this a bit.

Marking as critical, since it is not as big an overhead as the logging.