ksck against a large table appears to be reactor-bound on the client

Running ksck against a table with 36 tablets across 9 nodes (and ~6B rows of data), it seems like the ksck client is actually CPU bound in its reactor thread. (top -H shows the reactor at 100% CPU).

An easy fix might be to have it start more reactors, but we should also see why it's doing so many RPCs/second that it actually bounds the reactor like this.