Raise the per-server recursive-clients ceiling in fetchlimit
With the resolver now legitimately escalating to TCP after repeated
UDP timeouts to the same authoritative, each lame-server lookup
takes ~50% longer to fail. The recursive-client backlog therefore
peaks a little higher before the fetches-per-server auto-tune drops
the quota below 200.
Bump the upper bound for the burst-against-lame-server and recovery
steps from 200 to 250 to absorb that extra latency. The lower bound
and the final post-recovery target (clients <= 20) are unchanged.