vt: make Emulator.closed atomic to fix the Close/Read data race

[cameronsjo]

Closes #930

Emulator.closed is a plain bool that Close writes and Read reads with no synchronization between them, so -race flags every shutdown of an Emulator whose output is being consumed. Read blocks on an empty pipe and Close is the only thing that unblocks it, so there is no usage pattern that avoids the pair running concurrently.

SafeEmulator does not cover it either: its Read takes no lock (deliberately, so a parked reader can't hold mu against every writer) and it does not override Close, so the embedded *Emulator.Close is promoted unchanged.

The change

closed becomes an atomic.Bool. The guard is an early-out rather than the mechanism — the actual unblocking is pw.CloseWithError(io.EOF) on an io.Pipe, which is internally synchronized and idempotent — so no mutex is needed and Read stays lock-free.

Close uses CompareAndSwap rather than a Load then a Store, which also makes it single-winner: two concurrent Close calls previously could both pass the guard and both call CloseWithError. Harmless, since that call is idempotent, but the CAS form costs nothing and says what it means.

Verification

• go test -race ./... in vt/ passes.
• A reproducer that reliably fails on main passes with this change:

func TestCloseRacesParkedRead(t *testing.T) {
	e := vt.NewEmulator(80, 24)

	done := make(chan struct{})
	go func() {
		defer close(done)
		buf := make([]byte, 64)
		for {
			if _, err := e.Read(buf); err != nil {
				return
			}
		}
	}()

	time.Sleep(100 * time.Millisecond) // let the reader park inside Read
	if err := e.Close(); err != nil {
		t.Fatalf("Close: %v", err)
	}
	<-done
}

On main this reports the race at emulator.go:252 (Read) against emulator.go:265 (Close). With this change it passes.

• Downstream check: a terminal-multiplexing consumer of vt that had to exclude every shutdown-touching test from its -race job runs its full suite green, with no exclusions, against a patched vt.

I did not add the reproducer as a test in this PR since it needs a sleep to set up the condition; happy to include it in whatever form suits the package if you'd like it covered.

[cameronsjo]

Closing this — #881 already proposes the same atomic.Bool change and does it better, using Swap(true) for the close-once guard and adding a regression test this PR lacks. I missed it before opening; apologies for the noise.

Moving the verification detail over to #881.