Buffered events are lost when the process ends before a flush or tick #32

Adding an event or metric to the buffer wrote it to Redis immediately (persist-on-enqueue). That put a Redis round trip, up to the command timeout (5s by default) while Redis is slow, on the request path of every host call to sendEvent() / sendMetric().

The first fix on this branch moved the write off the request path: EventBuffer::add* only marks items pending, and persistPending() writes them at the end of flushBuffer() and from tick(). The trade-off is that a request that ends before any flush or tick never wrote its items, so they were lost.

I am closing that gap wherever PHP allows. The agent registers one shutdown function when Redis persistence is attached, and it calls persistPending(). That covers normal end, exit(), an uncaught exception, a fatal error, out of memory and max_execution_time. The request path still has no Redis round trip.

What cannot be fixed: SIGKILL (the OOM killer, kill -9) and SIGTERM without a host handler run no PHP code at all. I will prove all of this with child-process tests against a real Redis (bin/test_shutdown.php) and document the two unfixable cases with the reason.

I deliberately do not call fastcgi_finish_request() from the shutdown function (it would cut off output and shutdown functions the host registers later) and I do not install a SIGTERM handler (it would override the host's).