Wait for streaming responses to drain

[tnsardesai]

problem

Large streaming handlers can return while response bytes remain unacknowledged in the guest TCP send queue. Releasing the request scale-to-zero hold at that point can suspend the guest before the peer receives the response tail.

change

• wrap accepted TCP connections so response drains run after net/http writes final response framing, including the terminal chunk for chunked responses
• retain an additional request-counted scale-to-zero hold from handler completion through response finalization and TCP acknowledgement
• apply a five-minute write deadline before net/http finalizes the response, then poll Linux TIOCOUTQ every 100 ms within the same overall deadline
• release on acknowledgement, timeout, write-deadline failure, or socket-state failure, with structured outcome logs
• preserve current behavior for non-TCP listeners and routes that do not stream files or binary data

TIOCOUTQ support and its unsent-plus-unacknowledged semantics were validated in the current browser runtime.

tests

• go build ./...
• go vet ./...
• go test -race ./lib/scaletozero ./cmd/api/api
• go test -race $(go list ./... | grep -v '/e2e$')
• content-length slow-reader, chunked finalization, ceiling, immediate-drain, unsupported-deadline, ioctl-error, and non-TCP fallback coverage
• slow-reader and chunked/ceiling tests repeated 20 times under the race detector