Try reducing num-turbotunnel in server

#40200 (closed) (and later #40246 (closed)) caused snowflake-server to use multiple parallel KCP engines, which had an observable beneficial effect on performance. But those changes happened during the time between #40199 (closed) and #40262 (closed), and it may be that their apparent benefit was actually an accidental partial mitigation of #40260 (closed). (Partitioning clients over multiple KCP engines may have reduced the chance of packet corruption, which had its own negative effect on performance.)

I want to try reducing num-turbotunnel again, now that #40260 (closed) is fixed.

/cc @linus