Skip request body drain when connection will close

[emreay-]

Summary

Skip draining an unconsumed framed request body when the response has already committed the connection to close.

The post-response drain prevents unread body bytes from being parsed as the next request on a keep-alive connection. When the response contains Connection: close, there is no next request to protect.

Problem

If a ContentReader aborts an unterminated chunked upload, the current drain waits for the terminal chunk after sending the response. A client that keeps producing chunks continually resets the read timeout, so:

• the announced connection close never occurs;
• the client receives no transport-level failure signal; and
• a server worker remains occupied discarding data.

This also means the server can send Connection: close while continuing to consume the request indefinitely.

Fix

Use the finalized response header as the close decision:

if (!req.body_consumed_ && detail::has_framed_body(req)) {
  if (res.get_header_value("Connection") == "close") {
    connection_closed = true;
  } else {
    // Drain before reusing the connection.
  }
}

write_response_core already sets this header for keep-alive exhaustion, request-directed closure, handler-directed closure, and error responses. Reusable connections retain the existing drain behavior.

Relationship to #2450

This is a follow-up to #2450:

• Server closes keep-alive connection after DELETE request without Content-Length (drain bug) #2450 prevented the drain from reading an unframed request, where it could consume the next request.
• This change prevents the drain from running when the connection will close, where no next request exists.

Both changes restrict the drain to cases where it can serve its request-smuggling-prevention purpose.

Test

A raw-socket regression test sends an unterminated chunked request whose ContentReader rejects the first chunk. It verifies that the client receives the 409 Connection: close response followed by EOF.

On unpatched master, the final recv times out because the server remains in the drain. With this change, it returns 0 immediately.

The new test and the existing #2450 regression both pass on Windows with MSVC.

[yhirose]

@emreay- thanks for this PR. Could you please take a look at the test/style-check error?

[yhirose]

@emreay- thanks for this fine contribution!