qml: fix the Protocols.* bindings a manual QML corpus exercises

[jcelerier]

A corpus of manual QML protocol scenarios, run against real sockets from score's Mapper device, turns up a number of defects in the Protocols.* bindings. Each one below is covered by an integration test in the companion score PR, and each was proven by reverting the fix and watching that test go red.

Framing and transports

• Coalesced line framing dropped every line but the first. line_framing_decoder::receive() cleared the shared read buffer, but async_read_until reports the delimiter without consuming what follows it, so the rest of a coalesced read was destroyed. Now every complete line of a read is dispatched in order and a trailing partial line is kept for the next one.
• Line-framed sockets never observed EOF: the completion handler returned on ec.failed() before validate_stream, so the framing most hardware protocols use could not report a close. EOF now routes through validate_stream without re-arming a dead socket.
• An Intel HEX / S-record EOF record was dispatched as an empty message, overwriting the last decoded payload and echoing an empty frame back.
• A client never saw the peer close unless the script had a message callback: the read loop was only armed then. It is now always armed (payloads still only dispatched when a callback exists), an abortive close is reported like a graceful one, and on_close is emitted at most once.
• onFail was dead. The documentation block advertises it; every factory read only onError. It is now a documented alias.

WebSocket

• inboundWS did not listen: it allocated a state object, called onOpen and returned. It now runs a real server on the shared asio context, with a per-connection object, and releases the listener on destruction.
• The WebSocket client invoked its QJSValue callbacks straight from the websocketpp thread; they now hop to the object's Qt thread like everything else.

HTTP

There were two client implementations; the legacy Protocols.http(url, cb, verb) used the older one, which dropped the query string, sent a hard-coded Host, silently discarded every non-2xx response and had a no-op error path. It is now an adapter over the single implementation, which additionally learned: bodiless responses (HEAD, 1xx, 204, 304, 2xx-to-CONNECT) complete at end of headers, a body truncated before Content-Length is an error rather than a success, Content-Length is an upper bound, and the authority is built correctly (port when non-default, brackets for IPv6, punycode for IDN, userinfo, CRLF rejected in header values).

MIDI / UMP

port_information::api was not serialized, so an endpoint that round-tripped through QML came back as API(0) and libremidi refused to open it. It now round-trips, and an endpoint without a usable API reports why.

Mapper scripting

• An unqualified Device.write("/leaf") resolved against every device in the document and could land on a sibling; it now resolves in the script's own device, while name:/path still reaches any.
• Sockets opened during QML property evaluation ran onOpen before the Mapper's tree existed, so the Device.write they made was silently dropped. Opening is deferred to the event loop.

Build

Every target that compiles boost.asio in one binary must agree on BOOST_ASIO_ENABLE_BUFFER_DEBUGGING; libremidi did not, which is an ODR violation in Debug and shows up under ASan as a new-delete-type-mismatch on the resolver thread.