Add AllReduce collective (rapidsai#683)

Implements an `AllReduce` collective on top of the `AllGather` implementation. This may or may not be an acceptable long-term implementation given the performance characteristics it delivers, see the `AllReduce` class docs for details, but it makes the implementation certainly easier to maintain.

```
/**
 * @brief AllReduce collective.
 *
 * The current implementation is built using `coll::AllGather` and performs
 * the reduction locally after allgather completes. Considering `R` is the number of
 * ranks, and `N` is the number of bytes of data, per rank this incurs `O(R * N)` bytes of
 * memory consumption and `O(R)` communication operations.
 *
 * Semantics:
 *  - Each rank calls `insert` exactly once to contribute data to the reduction.
 *  - Once all ranks call `insert`, `wait_and_extract` returns the
 *    globally-reduced `PackedData`.
 *
 * The actual reduction is implemented via a type-erased `ReduceOperator` that is
 * supplied at construction time. Helper factories such as
 * `detail::make_reduce_operator` (defaults to host-side) or
 * `detail::make_device_reduce_operator` (device-side) can be used to build
 * element-wise reductions over contiguous arrays.
 */
```

Both host-side and device-side reductions are supported, whether it happens on host or device depends on the operator being passed, for built-in operators there are `HostReduceOp` and `DeviceReduceOp`, and custom operators require defining a `ReduceOperator` with `ReduceOperatorType::Host` or `ReduceOperatorType::Device` to determine where it runs. Finally, the implementation ensures data is properly located before executing the reduction, any device data is moved to host when running a host reduction, and host data is moved to device when running a device reduction.

Authors:
  - Peter Andreas Entschev (https://github.com/pentschev)
  - Bradley Dice (https://github.com/bdice)
  - Niranda Perera (https://github.com/nirandaperera)

Approvers:
  - Kyle Edwards (https://github.com/KyleFromNVIDIA)
  - James Lamb (https://github.com/jameslamb)
  - Lawrence Mitchell (https://github.com/wence-)
  - Mads R. B. Kristensen (https://github.com/madsbk)

URL: rapidsai#683

Author: pentschev

cpp/CMakeLists.txt

@@ -167,7 +167,8 @@ target_link_options(maybe_asan INTERFACE "$<$<BOOL:${RAPIDSMPF_ASAN}>:-fsanitize
 
 add_library(
   rapidsmpf
-  src/allgather/allgather.cpp
+  src/coll/allgather.cpp
+  src/coll/allreduce.cpp
   src/bootstrap/bootstrap.cpp
   src/bootstrap/file_backend.cpp
   src/bootstrap/utils.cpp

…nclude/rapidsmpf/allgather/allgather.hpp cpp/include/rapidsmpf/coll/allgather.hppcpp/include/rapidsmpf/allgather/allgather.hpp renamed to cpp/include/rapidsmpf/coll/allgather.hpp cpp/include/rapidsmpf/allgather/allgather.hpp renamed to cpp/include/rapidsmpf/coll/allgather.hpp

@@ -27,13 +27,13 @@
 #include <rapidsmpf/statistics.hpp>
 
 /**
- * @namespace rapidsmpf::allgather
- * @brief Allgather communication interfaces.
+ * @namespace rapidsmpf::coll
+ * @brief Collective communication interfaces.
  *
  * An allgather service for distributed communication where all ranks collect
  * data from all other ranks.
  */
-namespace rapidsmpf::allgather {
+namespace rapidsmpf::coll {
 namespace detail {
 
 /// @brief Type alias for chunk identifiers.
@@ -557,4 +557,4 @@ class AllGather {
     std::vector<std::unique_ptr<Communicator::Future>> receive_futures_{};
 };
 
-}  // namespace rapidsmpf::allgather
+}  // namespace rapidsmpf::coll