The existing implementation is now referred to as 'grouped-naive'.
It is pretty bad, because it groups together requests that need
to be dispatched to multiple servers.
'concurrent' is a new naive strategy, that is easy to implement
and should perform nicely.
'grouped-pk-serial' and 'grouped-pk-concurrent' still group
the ids they request, but in a smarter way, so each request group
only needs to access a single server.
I expect 'grouped-pk-concurrent' to be faster than 'grouped-pk-serial',
and it *may* be faster than 'concurrent' but we need benchmarks to know.