rpc: avoid ForkJoinPool compensation in send polling loop

[jkschneider]

What's changed?

RewriteRpc.send's polling loop now uses future.getNow(null) + Thread.sleep(1ms) instead of future.get(checkIntervalMs, TimeUnit.MILLISECONDS). Liveness check decoupled from polling cadence and fires every 500ms.

What's your motivation?

future.get(long, TimeUnit) from a ForkJoinPool worker goes through CompletableFuture.Signaller.block, which is a ForkJoinPool.ManagedBlocker. ManagedBlocker is the explicit hook FJP watches for and reacts to by spawning a compensation worker — a fresh thread added to keep parallelism while the original is parked.

That compensation worker can pick up other queued work (recipe load, printer invocations, etc.) and call PythonRewriteRpc.getOrStart() / JavaScriptRewriteRpc.getOrStart() / etc., spawning a fresh OS rpc subprocess into its ThreadLocal. When FJP later terminates the idle compensation worker, its TL is GC'd — but the OS process is independent of the JVM heap and survives. The dispatching worker's RunTask.execute() finally only runs shutdownCurrent() on the dispatching thread's TL, never on the dead compensation worker.

On a moderne-cli mod run against ~448 Python repos, this accumulated 127+ alive python rpc processes per long-running JVM, with each leaked rpc carrying a different past repo's log path in argv. Each compensation-worker spawn = one leaked OS process. The same mechanism applies to JS / C# / Go rpcs.

Thread.sleep parks the thread via LockSupport directly, not through ManagedBlocker, so FJP doesn't compensate. Parallelism temporarily drops by one while a worker is in rpc.send, which is exactly the resource-bounded behavior --parallel is supposed to mean.

Verified locally with a 20-repo Python LST sample running UpgradeToPython314 at --parallel=14:

• Before: spawn count >> repo count under load (compensation amplification)

• After: spawn count = repo count, alive rpc count ≤ --parallel, no compensation worker thread names in jstack, final_alive=0 after run.

• Pairs with rpc: kill child processes at JVM exit via shutdown hook #7616

• rpc: kill child processes at JVM exit via shutdown hook #7616 added a JVM-shutdown-hook on RewriteRpcProcess to kill child processes at JVM exit — that catches survivors at process termination. This PR prevents the leak from happening intra-run by removing its trigger.

Checklist

• I've read and applied the recipe conventions and best practices
• I've used the IntelliJ IDEA auto-formatter on affected files