skip invalid fields in write_headers to prevent response splitting

[metsw24-max]

The server serialises res.headers straight through write_headers, which formats each entry as name: value followed by CRLF without checking either part. The client side already guards this: req.headers go through check_and_write_headers, and the request line and websocket handshake run the same is_field_value check, so write_headers was the one output path left unguarded. Response::headers is a public field, and an application that reflects a request-derived value into it (a fairly common pattern) can carry a CR/LF into the response, splitting the header block and injecting further headers or a body. I noticed the gap while reading the recent chunked-trailer fix, which closed the same class for trailers but did not reach the ordinary header writer. The change skips any field whose name or value fails validation, which keeps the behaviour consistent with set_header rather than dropping the whole response. Fields the library emits itself all go through set_header and stay valid, so only directly-injected ones are affected. The regression test drives a handler that plants CR/LF in both a header name and value via res.headers and checks the split content never reaches the wire.

[yhirose]

@metsw24-max thanks!