Relacy tests

Instructions:

git clone -b stdexec https://github.com/ccotter/relacy && cd relacy
git clone -b relacy https://github.com/ccotter/stdexec
CXX_STD=c++20 make build/test/stdexec/{async_scope_bug,split_bug}
./build/test/stdexec/split_bug

Author: ccotter

include/exec/__detail/__bwos_lifo_queue.hpp

@@ -152,7 +152,7 @@ namespace exec::bwos {
       block_type(block_size, allocator),
       allocator_of_t<block_type>(allocator))
     , mask_(blocks_.size() - 1) {
-    blocks_[owner_block_].reclaim();
+    blocks_[owner_block_.load()].reclaim();
   }
 
   template <class Tp, class Allocator>
@@ -468,4 +468,4 @@ namespace exec::bwos {
   auto lifo_queue<Tp, Allocator>::block_type::is_stealable() const noexcept -> bool {
     return steal_tail_.load(std::memory_order_acquire) != block_size();
   }
-} // namespace exec::bwos
+} // namespace exec::bwos

include/exec/async_scope.hpp

@@ -781,7 +781,7 @@ namespace exec {
         // after this, which means we can rely on its self-ownership to ensure
         // that it is eventually deleted
         stdexec::start(
-          *new __op_t{nest(static_cast<_Sender&&>(__sndr)), static_cast<_Env&&>(__env), &__impl_});
+          *(new __op_t{nest(static_cast<_Sender&&>(__sndr)), static_cast<_Env&&>(__env), &__impl_}));
       }
 
       template <__movable_value _Env = empty_env, sender_in<__env_t<_Env>> _Sender>

include/exec/start_now.hpp

@@ -71,7 +71,7 @@ namespace exec {
       }
 
       void __complete() noexcept {
-        if (--__pending_ == 0) {
+        if (__pending_.fetch_sub(1) == 1) {
           auto __joiner = __joiner_.exchange(nullptr);
           if (__joiner) {
             __joiner->join();

include/stdexec/__detail/__when_all.hpp

@@ -191,7 +191,7 @@ namespace stdexec {
 
       template <class _Receiver>
       void __arrive(_Receiver& __rcvr) noexcept {
-        if (0 == --__count_) {
+        if (1 == __count_.fetch_sub(1)) {
           __complete(__rcvr);
         }
       }
@@ -361,7 +361,7 @@ namespace stdexec {
         } else if constexpr (!__same_as<decltype(_State::__values_), __ignore>) {
           // We only need to bother recording the completion values
           // if we're not already in the "error" or "stopped" state.
-          if (__state.__state_ == __started) {
+          if (__state.__state_.load() == __started) {
             auto& __opt_values = __tup::get<__v<_Index>>(__state.__values_);
             using _Tuple = __decayed_tuple<_Args...>;
             static_assert(