Synchronise the endpoint stats between the capture and stats threads

[git-hulk]

The stats thread walked syn_tab and read every query_stats in it while
the capture thread was inserting endpoints and bumping counters, with
nothing between them. Replaying a 16000 packet capture while polling
the stats endpoint reports six data races under ThreadSanitizer and
aborts: in stats_incr, in create_stats_object and on srv->stopped.

• Add stats_lock to the sniffer, held by the capture thread around
  every syn_tab lookup, insert and counter update, and by the stats
  thread across the whole walk so a response cannot be counted half
  way through serialisation. The requests table stays lock free: only
  the capture thread touches it.

• srv->stopped is written by server_terminate, which also runs from
  the signal handler, and read by the stats thread's loop. It is now
  sig_atomic_t accessed through the compiler's atomic builtins.

• Skip the endpoint instead of dereferencing NULL when the calloc for
  a new query_stats fails.

make test-race rebuilds under ThreadSanitizer and replays a generated
stress capture while polling the endpoint; it reports 0 races where the
previous code reported 6. It is opt-in because it clobbers the ordinary
build, and runs as its own CI job.

🤖 Generated with Claude Code