fix(client-cohttp-eio): don't block process exit on shutdown

[dijkstracula]

What

On a traced run (--trace) whose OTel collector is slow or unreachable, process
exit stalled for several seconds. This removes both causes.

1. Drop the 3s export backoff. send_http did Eio_unix.sleep 3. after a
   failed POST. Because cleanup's forced flush also goes through send_http,
   that sleep ran at teardown and added ~3s to every exit. Removed; the batch
   machinery (batch_timeout, drain-on-pop) already paces retries, and the
   error report is gated on not (Atomic.get stop) so a failed shutdown export
   stays quiet rather than spamming.

2. Daemonize the ticker. The 0.5s inter-tick Eio.Time.sleep was in a plain
   awaited fiber, so the switch waited it out at teardown (≤0.5s tail). It's now
   a fork_daemon, so the switch cancels the sleep at teardown. Safe because
   cleanup force-flushes the final batch before the switch returns.

Evidence

Fragment-scanner APM traces: Pro_core_CLI.main_exn is unchanged (~1.8s), while
the total dropped from ~5.9s (3s-sleep) to ~3.2s (no-sleep) — the delta was pure
shutdown tail.

On a traced run (--trace) when the OTel collector is slow or unreachable, the process exit took several seconds. The changes below fix both reasons.

1. Remove the 3‑second export backoff.
   send_http used to call Eio_unix.sleep 3. after a failed POST. Because cleanup also calls send_http, that sleep happened during teardown and added about 3 seconds to every exit. It has been removed. The batch system (batch_timeout, drain‑on‑pop) already controls retry timing, and the error report is only sent when not (Atomic.get stop) so a failed shutdown export stays silent instead of spamming.

2. Run the ticker as a daemon.
   The 0.5‑second pause between ticks used to be in a normal awaited fiber, so the switch waited for it during teardown (up to 0.5 seconds). Now it is started with fork_daemon, which lets the switch cancel the sleep at teardown. This is safe because cleanup force‑flushes the final batch before the switch returns.

Evidence

Fragment‑scanner APM traces show that Pro_core_CLI.main_exn time stayed about 1.8 seconds, while the total time dropped from roughly 5.9 seconds (with the 3‑second sleep) to about 3.2 seconds (without the sleep). The difference is just the shutdown tail.%