Description

Fixes:

• build.sh 的 build_device() 未转发 ENABLE_EXPERIMENTAL，导致 experimental BIRS 执行器未编入 device 库，运行时 ReduceScatter 失败
• ReduceScatter BIRS 多层忽略通信返回值导致失败后仍继续执行

Build-system changes required for the first fix:

• Third-party dependency pin bump: cann-cmake master-044 -> master-051 in cmake/fetch_cann_cmake.cmake
  (URL_HASH 2d8827b3... -> f50ea0bf...). Reason: only master-051's add_cann_device_project supports
  extra-argument passthrough (cmake_parse_arguments(CANN "" "" "EXT_ARGS" ${ARGN}) +
  list(APPEND EP_CMAKE_ARGS ${CANN_EXT_ARGS}) in function/prepare.cmake); master-044 has no EXT_ARGS.
  This is the prerequisite for CMakeLists.txt: add_cann_device_project(hccl EXT_ARGS -D ENABLE_EXPERIMENTAL=${ENABLE_EXPERIMENTAL}),
  i.e. the flag reaches the device ExternalProject without keeping an in-repo override copy of add_cann_device_project.
  Pin verified: SHA256 of https://raw.gitcode.com/cann/cmake/archive/refs/heads/master-051.tar.gz
  == f50ea0bf7f80ce9e664f26693252aee8d2a42438a80bc2cd335741dc4b3b834f.
  044 -> 051 side effects reviewed (new __cann_check_build_env() gate: gcc >= 7.3.0 + python3;
  convert_dependencies_to_package_formats renamed to a private __cann_ helper; device EP no longer forwards
  -D VERSION_INFO): hccl does not call the renamed function, and cmake/device/CMakeLists.txt includes
  version.cmake directly, so host and device builds are unaffected.
• cmake/device/CMakeLists.txt: added option(ENABLE_EXPERIMENTAL "Enable experimental features" OFF) so the
  standalone device project accepts the forwarded flag and gates add_subdirectory(experimental/ops).
• Removed the now-redundant if(NOT ENABLE_EXPERIMENTAL) return() guards from 6 leaf CMakeLists under
  experimental/ops (op_common, op_common/template, op_common/topo, reduce_scatter,
  reduce_scatter/birs, reduce_scatter/birs/template), per review. Gating stays single-sourced:
  the guard in experimental/ops/CMakeLists.txt plus the two add_subdirectory call sites
  (root CMakeLists.txt for host, cmake/device/CMakeLists.txt for device); leaf target_sources(hccl ...)
  remains protected by if(TARGET hccl), so the device build (no hccl target) is unaffected.
• experimental/ops/all_reduce/example: GetAlgNetMeta() signature adapted to the current base-class
  interface (added const OpParam& param).

Related Issue

https://gitcode.com/cann/hccl/issues/324
https://gitcode.com/cann/hccl/issues/380

Testing

Testing with the hccl-checker tool via ST tests "/test/st/algorithm/testcase/reduce_scatter_testcase_a3.cc"
Single-node: ReduceScatterBirs was deployed on one Atlas A3 node with RankSize set between 4 and 16.
Example of ReduceScatterBIRS output from hccl_test for RankSize 16:

Documentation Update

• docs/zh/build/build.md and docs/en/build/build.md: third-party software table, cann-cmake row
  master-044 -> master-051. Download source switched from the cann-3rd OBS bucket
  (cmake-master-051.tar.gz is not published there — HTTP 403) to the raw.gitcode.com archive, whose
  SHA256 matches the pin declared in cmake/fetch_cann_cmake.cmake.

Type Label

• Bug fix
• New feature
• Performance optimization
• Documentation update
• Other, please describe: