A short TCP send fails a TLS record: std.tcp has no write that sends a whole buffer #3108

tcp.write, tcp.write_n and tcp.write_slice each make one send(2) and return its count, and that count can be short. Every socket std.net opens has a 30 s send timeout (std/net/aether_net.c, net_set_socket_timeouts). When a peer reads slowly enough that one send runs into that timeout, the send returns the bytes that fit by then, and a signal can cut a send short the same way. std.tcp has nothing that sends a whole buffer, so each caller has to loop on the rest.

std/cryptography/tls13_client and tls13_server don't. They send each record with one tcp.write_n, and a short count is reported as "failed to send record". By then the peer already holds part of the record, so the handshake or the connection is lost.

Fix:

• Add tcp.write_all(sock, data: byte[]). It sends until every byte is out or a send fails, and returns (sent, err).
• The TLS modules send their records through it.
• A test forces short sends (a small send buffer, a short send timeout, a peer that reads slowly). On a platform where one send comes back short, it shows that, and on every platform it checks that write_all delivers every byte in order.

contrib/tinyweb has the same problem in its WebSocket, SSE and chunked writes, along with frame bugs of its own; those are #3109.

Found on #3101 (#3107's migration of the TLS modules onto std.tcp).