feat: support ReduceScatter with OpenMPI backend implementation

[halfman510]

Summary

This PR introduces an OpenMPI-based implementation of ReduceScatter, along with a complete example program for functionality verification and basic performance evaluation.

Changes

• OpenMPI-based ReduceScatter Implementation
  • add the basic OpenMPI implementation for infiniReduceScatter(), including:
    • the core interface src/base/reduce_scatter.h;
    • the OpenMPI backend implementation in src/ompi/impl/reduce_scatter.h;
    • the public API declaration in include/infiniccl.h.
  • add an example program examples/reduce_scatter.cc similar to examples/all_reduce.cc for correctness verification and simple performance testing.

Known Issues & Future Work

• The current OpenMPI AllGather implementation uses blocking MPI_Reduce_scatter_block, which prevents overlap between communication and computation. Future work may introduce non-blocking collectives (MPI_Reduce_scatter_block) and stream-aware asynchronous execution to improve concurrency and performance.
• The current implementation allocates temporary host staging buffers using malloc/free on every invocation. This may introduce noticeable overhead in high-frequency workloads. Future work may add reusable buffer pools, allocator caching, and pinned host memory support to improve transfer efficiency and reduce allocation overhead.
• For the heterogeneous UCX + InfiniBand cluster used in testing, large AllGather messages (e.g., 1 << 20 elements) may fail with mlx5 RC RDMA_READ errors due to UCX rendezvous RDMA_READ path limitations. This requires setting UCX_RNDV_SCHEME=put_zcopy to force a safe put-based transfer protocol. Without this setting, large-message AllGather execution is unstable on some NIC configurations.
• Averaging (kAvg) is performed via a CPU-side loop after the MPI call. While functionally correct, this is not optimal for large recv_count. Future work may move scaling into the MPI operation (where supported) or use a more efficient vectorized/device-side post-processing step.
• recv_count is cast to int for MPI (with a safety check). Extremely large messages exceeding INT_MAX elements are rejected. This is acceptable for current use cases but may need MPI_Count support in future MPI-4+ integrations for very large tensors.

Logs & Screenshots

• all_reduce test (MetaX-NVIDIA heterogeneous)
  all_reduce.log

• all_gather test (MetaX-NVIDIA heterogeneous)
  all_gather.log

• reduce_scatter test (MetaX-NVIDIA heterogeneous)
  reduce_scatter.log