Skip the synchronous getaddrinfo for a hostname that can never resolve - #40986

What is and is not shown here:

• The blocking resolver is not observed in this PR. It is the same libc getaddrinfo call, on the same names, that dns: reject names that are not hostnames before any resolver, report empty dns_sd answers as ENOTFOUND #40970 measured on the connect side (dns.lookup waiting out the macOS resolver timeout for a label with a space). Linux glibc rejects these names locally, so on Linux the visible change is only the error.
• No user report exists for a malformed bind hostname. Bun.serve failed to start on localhost subdomain #25765 (something.localhost on macOS) is a well-formed name and is Report an unresolvable listen hostname as a getaddrinfo error #37690's case.

Repro on 1.4.1 (Linux, no resolver needed for the first four names):

serve "this is not a hostname" => ENOENT listen "ENOENT: no such file or directory, listen"
listen "this is not a hostname" => (no code) "Failed to listen at this is not a hostname"
udpSocket "this is not a hostname" => ENOENT bind "bind ENOENT this is not a hostname"
udpSocket connect { hostname: "this is not a hostname" } => DNSException "connect ENOTFOUND this is not a hostname"
dgram socket.connect(1234, "this is not a hostname") via a custom lookup => no error, socket marked connected
serve "aaaa...(64).com" => EMSGSIZE listen
serve "nxdomain.invalid" => EAGAIN listen (left to #37690)
udpSocket "nxdomain.invalid" => ESRCH bind (left to #37707)

After this change the first six report getaddrinfo ENOTFOUND <host> (name: "Error", code: "ENOTFOUND", syscall: "getaddrinfo", hostname). node:net, node:http and node:dgram (custom lookup) servers emit the same error object from listen() / bind().

Where the check sits:

• Bun.serve: at the top of NewServer::listen (src/runtime/server/mod.rs), before the uws app is created. The existing deinit path frees the server and Bun.serve throws synchronously, as it does for EADDRINUSE. The bracket strip now lives in one function, strip_ipv6_brackets, used by the check and by the existing Addr code. It used to drop the first and last byte of any hostname that starts with [, so [foo] resolved as foo ([localhost], [127.0.0.1] and [example.com were already rejected by the base URL parse). It now strips only when the inside is an IPv6 literal (to_ip_address, so a %zone still passes); anything else keeps its brackets and fails is_valid_hostname.
• Bun.listen (src/runtime/socket/Listener.rs): right after SocketConfig::from_js, only when a port is set (a unix path or an adopted fd has no hostname). SocketConfig::from_js itself is shared with Bun.connect, whose error is delivered asynchronously, so the check is not placed there.
• Bun.udpSocket (src/runtime/socket/udp_socket.rs): bind and connect.hostname are both checked right after the config parse, before the socket is created. Bind failures there throw synchronously today, and so does this. The connect option used to report these names as connect ENOTFOUND <host> (DNSException) from the EAI_* code; a well-formed unresolvable connect.hostname still does, that block is udp: report a failed connect() as the socket error, not a getaddrinfo error #38622's.
• The internal UDPSocket.jsConnect that node:dgram's socket.connect() calls gets the same check. dgram resolves with dns.lookup first, so only a custom lookup can hand it a raw name. glibc rejected such a name with a code jsConnect never read (it checks for -1 only), so the socket was marked connected with no error. udp: report a failed connect() as the socket error, not a getaddrinfo error #38622 reworks that return-code handling; the check here sits before the call and does not overlap it.

The error is built by not_a_hostname_error in src/runtime/dns_jsc/cares_jsc.rs from system_error_with_syscall_and_hostname, the helper Bun.connect uses.

For #37690 and #37707: their tests use a 64-byte label, which is_valid_hostname rejects, so once this lands that input is answered in-process and no longer exercises the dns_error plumbing. glibc rejects a*b.com, a/b.com, -a.com and ünicode.com locally (EAI_NONAME, no query sent) while is_valid_hostname accepts all four. Other platforms need checking in CI.

Bun.listen({ hostname: "[::1]" }) and Bun.udpSocket({ hostname: "[::1]" }) never stripped brackets and already fail on 1.4.1 (Failed to listen at [::1], bind ENOENT [::1]). They now fail with getaddrinfo ENOTFOUND [::1].

Suites run on the debug build: serve-listen.test.ts (37 pass), socket-dns-error.test.ts (11 pass), udp_socket.test.ts (214 pass), dgram.test.ts + udp_socket_recv_flags.test.ts (64 pass), serve.test.ts + serve-http3.test.ts + server-url-invalid.test.ts (349 pass, 2 environment failures: root can bind port 1003, egress proxy), node-net-server.test.ts, serve-epoll-add-fail.test.ts, test/internal/source-lints/ (171 pass), node test-dgram-bind*, test-dgram-connect*, test-dgram-custom-lookup, test-dgram-error-message-address, test-net-listen-error, test-http-listening. socket.test.ts and node-net.test.ts have the same localhost dual-stack failures on the released bun in this container.