[INFRA] Import NVIDIA/CCCL upstream as optimization reference library

CCCL (CUDA C++ Core Libraries) provides:
- CUB: device/block/warp-level GPU primitives (reduce, scan, sort, topk)
- Thrust: high-level parallel algorithms (transform_reduce, sort, scan)
- libcudacxx: CUDA C++ standard library (atomics, barriers, memory)
- cudax: experimental features (memory resources, allocators)
- Tuning policies: per-SM hardware-specific algorithm parameters

Competition optimization vectors mapped to CCCL:
- Output TPS (83% weight): warp_reduce, block_reduce, device_topk
- Input TPS (14% weight): device_scan, block_load, prefetch
- Cache TPS (3% weight): prefix caching strategy patterns
- Memory (0.9 util): pooled/cached/buddy allocators

Source: https://github.com/NVIDIA/cccl (shallow clone, HEAD only)
License: Apache-2.0

cccl_upstream/cudax/test/execution/common/impulse_scheduler.cuh

@@ -0,0 +1,206 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of CUDA Experimental in CUDA C++ Core Libraries,
+// under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+// SPDX-FileCopyrightText: Copyright (c) 2024 NVIDIA CORPORATION & AFFILIATES.
+//
+//===----------------------------------------------------------------------===//
+
+#pragma once
+
+#include <cuda/__utility/immovable.h>
+
+#include <cuda/experimental/execution.cuh>
+
+// IWYU pragma: begin_keep
+#include <condition_variable>
+#include <functional>
+#include <memory>
+#include <mutex>
+// IWYU pragma: end_keep
+
+#include "testing.cuh" // IWYU pragma: keep
+
+namespace ex = cuda::experimental::execution;
+
+#if _CCCL_HOST_COMPILATION()
+
+namespace
+{
+namespace _impulse
+{
+struct _attrs_t
+{
+  constexpr auto query(ex::get_completion_behavior_t) const noexcept
+  {
+    return ex::completion_behavior::asynchronous;
+  }
+};
+} // namespace _impulse
+
+//! Scheduler that will send impulses on user's request.
+//! One can obtain senders from this, connect them to receivers and start the operation states.
+//! Until the scheduler is told to start the next operation, the actions in the operation states are
+//! not executed. This is similar to a task scheduler, but it's single threaded. It has basic
+//! thread-safety to allow it to be run with `sync_wait` (which makes us not control when the
+//! operation_state object is created and started).
+struct impulse_scheduler : _impulse::_attrs_t
+{
+private:
+  //! Command type that can store the action of firing up a sender
+  using _cmd_t     = std::function<void()>;
+  using _cmd_vec_t = std::vector<_cmd_t>;
+
+  struct _data_t : std::enable_shared_from_this<_data_t>
+  {
+    explicit _data_t(int id)
+        : id_(id)
+    {}
+
+    int id_;
+    std::mutex mutex_{};
+    std::condition_variable cv_{};
+    std::vector<std::function<void()>> all_commands_{};
+  };
+
+  //! That data_t shared between the operation state and the actual scheduler
+  //! Shared pointer to allow the scheduler to be copied (not the best semantics, but it will do)
+  std::shared_ptr<_data_t> _data_{};
+
+  template <class Rcvr>
+  struct _opstate_t : cuda::__immovable
+  {
+    using operation_state_concept = ex::operation_state_t;
+
+    _data_t* _data_;
+    Rcvr _rcvr_;
+
+    explicit _opstate_t(_data_t* data, Rcvr&& rcvr)
+        : _data_(data)
+        , _rcvr_(static_cast<Rcvr&&>(rcvr))
+    {}
+
+    void start() noexcept
+    {
+      // Enqueue another command to the list of all commands
+      // The scheduler will start this, whenever start_next() is called
+      std::unique_lock lock{_data_->mutex_};
+      _data_->all_commands_.emplace_back([this]() {
+        if (ex::get_stop_token(ex::get_env(_rcvr_)).stop_requested())
+        {
+          ex::set_stopped(static_cast<Rcvr&&>(_rcvr_));
+        }
+        else
+        {
+          ex::set_value(static_cast<Rcvr&&>(_rcvr_));
+        }
+      });
+      _data_->cv_.notify_all();
+    }
+  };
+
+  struct _sndr_t
+  {
+    using sender_concept = ex::sender_t;
+
+    _data_t* _data_;
+
+    template <class Self>
+    _CCCL_HOST_DEVICE static constexpr auto get_completion_signatures()
+    {
+      return ex::completion_signatures<ex::set_value_t(), ex::set_stopped_t()>();
+    }
+
+    template <class Rcvr>
+    auto connect(Rcvr rcvr) -> _opstate_t<Rcvr>
+    {
+      return _opstate_t<Rcvr>{_data_, static_cast<Rcvr&&>(rcvr)};
+    }
+
+    auto get_env() const noexcept
+    {
+      return _impulse::_attrs_t{};
+    }
+  };
+
+  explicit impulse_scheduler(_data_t* data)
+      : _data_(data->shared_from_this())
+  {}
+
+public:
+  using scheduler_concept = ex::scheduler_t;
+
+  impulse_scheduler()
+      : _data_(std::make_shared<_data_t>(0))
+  {}
+
+  explicit impulse_scheduler(int id)
+      : _data_(std::make_shared<_data_t>(id))
+  {}
+
+  ~impulse_scheduler() = default;
+
+  //! Actually start the command from the last started operation_state
+  //! Returns immediately if no command registered (i.e., no operation state started)
+  bool try_start_next()
+  {
+    // Wait for a command that we can execute
+    std::unique_lock lock{_data_->mutex_};
+
+    // If there are no commands in the queue, return false
+    if (_data_->all_commands_.empty())
+    {
+      return false;
+    }
+
+    // Pop one command from the queue
+    auto cmd = std::move(_data_->all_commands_.front());
+    _data_->all_commands_.erase(_data_->all_commands_.begin());
+    // Exit the lock before executing the command
+    lock.unlock();
+    // Execute the command, i.e., send an impulse to the connected sender
+    cmd();
+    // Return true to signal that we started a command
+    return true;
+  }
+
+  //! Actually start the command from the last started operation_state
+  //! Blocks if no command registered (i.e., no operation state started)
+  void start_next()
+  {
+    // Wait for a command that we can execute
+    std::unique_lock lock{_data_->mutex_};
+    while (_data_->all_commands_.empty())
+    {
+      _data_->cv_.wait(lock);
+    }
+
+    // Pop one command from the queue
+    auto cmd = std::move(_data_->all_commands_.front());
+    _data_->all_commands_.erase(_data_->all_commands_.begin());
+    // Exit the lock before executing the command
+    lock.unlock();
+    // Execute the command, i.e., send an impulse to the connected sender
+    cmd();
+  }
+
+  _sndr_t schedule() const noexcept
+  {
+    return _sndr_t{_data_.get()};
+  }
+
+  friend bool operator==(const impulse_scheduler& a, const impulse_scheduler& b) noexcept
+  {
+    return a._data_ == b._data_;
+  }
+
+  friend bool operator!=(const impulse_scheduler& a, const impulse_scheduler& b) noexcept
+  {
+    return a._data_ != b._data_;
+  }
+};
+} // namespace
+
+#endif // _CCCL_HOST_COMPILATION()