Expire pending requests instead of keeping them for ever

[git-hulk]

Every request went into sniffer->requests and only came out when a
matching response arrived. Requests that were never answered -- a
dropped connection, a server that never replied, a capture that started
mid-response -- stayed for the life of the process, so a long running
capture leaked steadily. It also meant a response arriving minutes
later was still reported as a latency of minutes: replaying a capture
whose answer comes 99 seconds later printed 99000.000 ms.

• hashtable gains a size counter and hashtable_sweep, which removes
  every entry a predicate selects, correctly unlinking at the head,
  the middle or the tail of a bucket chain.

• requests_expire drops requests older than 60 seconds of capture
  time. The sweep runs at most once per captured second, and is
  driven by packet timestamps rather than the wall clock so that
  replaying a file behaves the same as a live capture. Timestamps that
  go backwards evict nothing.

• 60 seconds is past the last latency bucket, so anything still
  waiting is lost rather than useful. MAX_PENDING_REQUESTS is a firm
  ceiling for the case where requests arrive faster than they age out.

• hashtable_add now checks its allocations, and no longer half-inserts
  an entry whose key could not be duplicated.

🤖 Generated with Claude Code