node: make the isolate the port namespace for http and net servers

[guybedford]

Follow-on to #7306, which made the virtual port table per-DO, when instead it should be per-isolate.

#7306 gave each Durable Object instance its own virtual port table, with the isolate's table underneath it. That made the scope of a listen() depend on where it was called: at module top-level it bound in the isolate table, in a Durable Object constructor or handler it bound in that instance's table, in the same worker. It also broke the Vite and Vitest plugins, which evaluate user modules inside a runner Durable Object: a top-level server.listen(8080) landed in the runner instance's table and httpServerHandler from the stateless entrypoint could not find it (cloudflare/workers-sdk CI).

This fixes the regressed DO behavior, with the simpler per-isolate scoping model, which is also the correct model since the isolate is the VM abstraction not the DO:

• one tcpPorts table per isolate; IoContext::getPortScopeKey and SocketsModule::getPortScopeKey are removed
• one host address per isolate (240.1.0.1) rather than one per Durable Object instance; the gateway address is unchanged
• owners no longer capture the table they bound in: bindPort() drops its table parameter, BoundSocket, net.Server, net.Socket, and http.Server use tcpPorts directly, and kBoundTable becomes a kBoundReserved flag

Tests: testDurableObjectScoping becomes testDurableObjectsShareIsolatePorts, covering a net.Server listened on in one instance holding the port against another instance, the entrypoint, and a BoundSocket, being reached through stub.connect() on either instance with the isolate host address and gateway peer, reservations shared and released across instances and the entrypoint, and an http.Server listened on inside an instance served through httpServerHandler from the entrypoint and another instance, with the reverse direction and the not-found error after close.

[ask-bonk]

I'm Bonk, and I've done a quick review of your PR.

Makes Node HTTP/net ports isolate-scoped across Durable Objects.

1. P1: Posted one inline suggestion covering cross-IoContext Durable Object dispatch.

Time for a pun! This port scope needs an actor to keep it in context.

github run

[petebacondarwin]

This should resolve the Vite/Vitest regression, thanks!
Are there any concerns about ports being broken when a DO gets moved from one metal to another? I assume that it doesn't matter since the DO must have died (or is killed) and so any sockets will already get closed and need to be re-opened.