Random tonlib::EngineConsoleClient-adjacent changes

[DanShaders]

tonlib::EngineConsoleClient was accidentally merged with simplex even though the code was not meant to be seen by the world yet. This PR fixes stuff I was meaning to change before upstreaming the client.

[github-actions]

• High: Data race in ObjectCounter now that count_ is plain size_t with no synchronization. new_actor() increments and the guard decrements can run on different threads (client thread vs scheduler thread), making count_ and is_zero() UB under concurrent use. Consider restoring std::atomic<size_t> or guarding with a mutex. tonlib/tonlib/FFIEventLoop.h:11 tonlib/tonlib/FFIEventLoop.cpp:10
• Medium: The actor-lifetime tracking changed from being tied to the actor instance (via ClientWrapper holding the guard) to being tied to the owning FFIEngineConsoleClient object (counter_). If EngineConsoleClient destruction is asynchronous or the actor outlives the owner, FFIEventLoop can now see is_zero() and stop the scheduler while an actor is still alive. If the intent is still “no stray actors”, keep the guard owned by the actor or ensure ActorOwn destruction is synchronous. tonlib/tonlib/FFIEngineConsoleClient.cpp:5 tonlib/tonlib/FFIEngineConsoleClient.h:31
• Low: Tests now dereference *loop.wait(-1) directly. wait() can return nullopt on cancellation or if the eventfd wait returns without data (e.g., EINTR). This can make tests flaky with a hard crash instead of cleanly retrying. Consider keeping the loop‑until‑value helper in tests. test/tonlib/test-ffi-awaitable.cpp:104 test/tonlib/test-ffi-event-loop.cpp:51

[github-actions]

• Commit da43f8d: tonlib/tonlib/FFIEventLoop.h:15-26 + tonlib/tonlib/FFIEventLoop.cpp:10-13 — ObjectCounter::count_ is now a plain size_t with unsynchronized ++/-- and reads in is_zero(). new_actor() and guard destruction can run on different threads (e.g., FFI callers), and FFIEventLoop::~FFIEventLoop() reads it concurrently, which is a data race/UB and can also make the CHECK unreliable. At minimum keep std::atomic<size_t> (as before) or guard with a mutex; if you truly rely on single-threaded access, that invariant needs to be enforced/ documented and the CHECK should still be synchronized.

[github-actions]

1. Commit e74aa1a2 (“Simplify check for no stray actors on FFIEventLoop destruction”): FFIEngineConsoleClient now holds the guard in the outer object, but the actor is reset asynchronously in ~FFIEngineConsoleClient(). That means counter_ can be released before the actor actually stops, so FFIEventLoop::~FFIEventLoop() can see object_counter_ as zero while a live EngineConsoleClient actor still exists. This defeats the “no stray actors” check and can allow event loop teardown with active actors. Consider re-attaching the guard to the actor lifetime (restore the wrapper), or move counter_ into the run_in_context lambda and only release it after client.reset() runs on the scheduler thread. References: tonlib/tonlib/FFIEngineConsoleClient.cpp:7, tonlib/tonlib/FFIEngineConsoleClient.h:19, tonlib/tonlib/FFIEventLoop.cpp:11.