fix(reactor): clear the op keepalive before publishing its continuation

coro_resume published the completion's continuation through dispatch_coro
before moving impl_ptr out of the op. When the continuation targets a
strand drained on another thread, that thread can resume it and reuse the
op via reset() the instant it is published, racing the impl_ptr move on
the completing thread. Move the keepalive out first so the write is
ordered before the publish; the ref still lives in the local across
resume(), so the op storage survives the continuation.

Author: sgerbino

include/boost/corosio/native/detail/coro_op_complete.hpp

@@ -125,8 +125,11 @@ inline void
 coro_resume(coro_op* self) noexcept
 {
     self->cont.h = self->h;
-    auto next = dispatch_coro(self->ex, self->cont);
+    // Clear the keepalive before publishing the continuation: a strand
+    // drained on another thread can reuse this op via reset() the instant
+    // it runs, so this write must be ordered before the publish, not after.
     auto suicide = std::move(self->impl_ptr);
+    auto next    = dispatch_coro(self->ex, self->cont);
     next.resume();
     // suicide drops here; may destroy impl + self.
 }