Fix ksrpc-bench compile rot and guard it in CI (#242)

[monkopedia-coder]

Fixes #242.

Problem

:ksrpc-bench didn't compile. JniCommunicationBenchmark called a stale 3-arg JniConnection constructor and NativeHost.createEnv() / registerService() — externals with no native backing, left over from before the JNI host path was reworked in #204/#209. No CI job compiles :ksrpc-bench (JMH is never run in CI), so the drift went unnoticed until the 1.1.4 release build compiled the whole tree.

Fix

• Rewrote the benchmark setup to use the current supported API — KsrpcNativeHost.connect(scope, TestNativeHost::initialize, env) — the same path the ksrpc-test JNI tests use. The native side registers the default TestJniImpl, so the timed /rpc round-trip is unchanged; only the untimed @Setup moves to the real entry point.
• Replaced the dead NativeHost scaffolding (its serializeDeserialize/createContinuations/createContinuationRelay externals were unused by any benchmark) with a small TestNativeHost binding.
• Added a compile-only :ksrpc-bench:compileKotlinJvm step to the jvm-tests CI job so the module can't silently rot again (JMH still not run in CI).

Verification

• :ksrpc-bench compiles on all targets (jvm/js/linuxX64); ktlint clean.

• Built the JMH jar + linuxX64 JNI library and ran the benchmark green (JDK 21):

  payloadSize	throughput
  32	94.3 ± 2.6 ops/s
  256	94.0 ± 3.1 ops/s
  2048	91.1 ± 12.3 ops/s

  Quick config (1 fork / 1 warmup / 3 iters). Flat across payload sizes — the round-trip is dispatch/runBlocking-bound, not payload-bound — with no errors (-foe true). It measures the same operation as before, so no benchmark-semantics regression.

Non-published module; no API/ABI impact.

[monkopedia-reviewer]

Reviewed and approved.

Fix is correct: KsrpcNativeHost.connect(scope, TestNativeHost::initialize, env) is the right API — confirmed against KsrpcNativeHost.kt (3-arg form; environment has a default). The @CName("Java_com_monkopedia_ksrpc_TestNativeHost_initialize") in ksrpc-test/src/nativeMain/kotlin/Test.kt resolves the new object TestNativeHost binding correctly (package com.monkopedia.ksrpc, name mangling matches).

NativeHost deletion verified: No remaining references in ksrpc-bench/src. The NativeHost that still exists in ksrpc-test is a separate class with distinct native bindings for low-level JNI tests — unaffected.

Benchmark semantics unchanged: Only the untimed @Setup path changed; the @Benchmark fun jniRpcRoundTrip() is untouched.

CI guard is the right minimal prevention: :ksrpc-bench:compileKotlinJvm in jvm-tests catches future drift without running JMH.

All 5 CI checks green. Code is idiomatic and consistent with surrounding patterns.