Blocking sends to stop channel can deadlock goroutines

All three middleware packages (http, cloudevents, kafka) use a buffered channel `stop chan error` (buffer size 1) for shutdown signaling. Multiple goroutines send to this channel (HTTP server, consumer/client loop, Start hook, signal handler), but only one value is ever read by the `select` in `Start()`.

If the buffer is already full when another goroutine tries to send, or if `Start()` has already returned and no receiver remains, the sending goroutine blocks forever (leaked goroutine).

**Fix:** Replace blocking sends with non-blocking sends:

```go
select {
case s.stop <- err:
default:
}
```

**Affected files:**
- `http/service.go`
- `cloudevents/service.go`
- `kafka/service.go`

Found during review of #169.

Start with the blocking stop-channel sends in http/service.go, cloudevents/service.go, and kafka/service.go, comparing how each middleware signals shutdown. Replace the blocking sends in all three files with the specified non-blocking behavior, then verify that shutdown signaling no longer leaves goroutines blocked.

Written by the indexing model from the issue text.