fix(runtime): native poll(2) for guest stdio + event-driven kernel waits

[NathanFlurry]

• The wasm runner's net_poll (the patched libc's poll(2) backend) now reports readiness for every fd type — kernel-managed stdio/PTY fds via __kernel_poll, pipes, regular files, and POLLNVAL for closed fds — instead of host_net sockets only, and sleeps between iterations instead of hot-spinning. select()-based guests (vim) now see stdin readiness and no longer burn CPU while idle.
• The sidecar services __kernel_stdin_read / __kernel_poll without blocking the dispatch loop: a zero-timeout readiness probe answers immediately when possible, otherwise the RPC is parked (reply-by-token) and an off-loop waiter on the kernel's poll notifier re-enqueues it when poll state changes. Child processes park on the child-event pump with per-iteration rechecks.
• Runner input waits switched from 10ms polling slices to long event-driven waits (host_tty.read, readKernelStdinChunk, poll_oneoff, net_poll), so idle full-screen guests accrue ~zero active CPU.
• The wasm "fuel" wall-clock timeout is now opt-in (explicit max_fuel only). Runaway modules stay bounded by the default 30s active-CPU V8 watchdog, while idle interactive guests are no longer killed on wall time.
• The kernel exposes a cloneable PollWaitHandle for waiting on poll-state changes off the kernel owner's thread.
• Kill/teardown flushes a parked kernel-wait RPC (EINTR) first, since isolate termination cannot interrupt the native bridge wait — prevents teardown deadlocking against a blocked guest thread.