[Core][MPI] Add non-blocking (asynchronous) operations to DataCommunicator

[loumalouomega]

📝 Description

DataCommunicator only supported blocking calls. This PR adds non-blocking point-to-point and collective operations. They return a new MPI-agnostic DataCommunicatorRequest, which is the core's equivalent of MPI_Request.

Key changes

• DataCommunicatorRequest (kratos/includes/data_communicator_request.h): a move-only handle with Wait(), Test(), IsCompleted() and a static WaitAll(). A default-constructed request counts as already completed. Destroying a pending request waits for it to finish.
• New DataCommunicator methods:
  • ISend / IRecv (MPI_Isend / MPI_Irecv)
  • IBarrier (MPI_Ibarrier)
  • IBroadcast (MPI_Ibcast)
  • ISumAll / IMinAll / IMaxAll (MPI_Iallreduce)
  • Types: the same as Send/Recv (char, int, unsigned int, long unsigned int, double, array_1d, Vector, Matrix, plus std::vector of each), and std::string for send/recv/broadcast.
• MPIDataCommunicator: the request owns the MPI_Request and the message buffers. Received data that isn't contiguous (e.g. std::vector<Vector>) is copied back on completion.
• Serial DataCommunicator: operations complete immediately; IRecv throws, like Recv.
• Python: int/double variants plus ISendString, IRecvString and IBarrier. They return a DataCommunicatorRequest that keeps the buffers alive; read the result with GetResult().

Usage rules: output buffers must be sized in advance, including the inner shapes of Vector/Matrix entries, because no size information is sent. Pair ISend with IRecv.

std::vector<double> recv(n);
std::vector<DataCommunicatorRequest> requests;
requests.push_back(r_comm.IRecv(recv, prev_rank, tag));
requests.push_back(r_comm.ISend(send, next_rank, tag));
// ... overlap computation ...
DataCommunicatorRequest::WaitAll(requests);

double global_sum = 0.0;
auto request = r_comm.ISumAll(local_value, global_sum);
// ...
request.Wait();

req = comm.IRecvDoubles(2, recv_rank)
comm.ISendDoubles([1.0, 2.0], send_rank).Wait()
req.Wait()
values = req.GetResult()

total = comm.ISumAll(rank)
KM.DataCommunicatorRequest.WaitAll([total])
print(total.GetResult())

Validation

New tests:

• kratos/tests/cpp_tests/includes/test_data_communicator_request.cpp (serial)
• kratos/mpi/tests/cpp_tests/sources/test_mpi_data_communicator_async.cpp (MPI, ring exchange, all types, Test() polling, WaitAll, completion on destruction, collectives)
• testAsync* cases in kratos/mpi/tests/test_mpi_data_communicator_python.py

Results:

• Serial and MPI tests pass with mpirun -np 1/2/3.
• The full KratosMPICoreTest suite has no failures.

🆕 Changelog

• Added DataCommunicatorRequest, an MPI-agnostic handle for non-blocking communication
• Added ISend, IRecv, IBarrier, IBroadcast, ISumAll, IMinAll and IMaxAll to DataCommunicator / MPIDataCommunicator
• Exposed the non-blocking operations to Python