feat(nccl): add all-gather communication

[voltjia]

Summary

Implement the existing infinicclAllGather API for the NCCL backend and add the corresponding example programs.

Changes

• Support NCCL AllGather

  • Map the generic AllGather operation directly to ncclAllGather while preserving NCCL's per-rank count and legal in-place buffer contract.

• Examples

  • Add examples/ccl/all_gather.cc, a deterministic single-node multi-GPU example using one worker thread per GPU;
  • Add ·examples/ccl_mpi_hybrid/all_gather.cc·, which performs all_gather between global MPI-launched ranks;
  • Make the existing MPI examples/mpi/all_gather.cc example propagate validation failures through its exit status.

Platform and Backend Affected

Platform

• CPU
• NVIDIA GPU
• Iluvatar GPU
• MetaX GPU
• Moore Threads GPU
• Cambricon MLU
• HYGON DCU

Backend

• OpenMPI
• MPICH
• NCCL
• MCCL

Performance Impact

• No performance impact
• Performance improved
• Performance regression possible

N/A. This PR enables an existing API for NCCL and makes no performance claim.

Known Issues & Future Work

• N/A.

Test Results

Test Involved Platform

• CPU
• NVIDIA GPU
• Iluvatar GPU
• MetaX GPU
• Moore Threads GPU
• Cambricon MLU
• HYGON DCU

Test Involved Backend

• OpenMPI
• MPICH
• NCCL
• MCCL

NVIDIA NCCL + OMPI (Single Node):
ccl_mpi_hybrid_all_gather.log
ccl_mpi_hybrid_all_reduce.log
ccl_mpi_hybrid_send_recv.log

NVIDIA NCCL (Single Node):
ccl_all_gather.log
ccl_all_reduce.log
ccl_send_recv.log

Iluvatar NCCL + OMPI (Single Node):
ccl_mpi_hybrid_all_gather.log
ccl_mpi_hybrid_all_reduce.log
ccl_mpi_hybrid_send_recv.log

HYGON RCCL + OMPI (2 Nodes):
ccl_mpi_hybrid_all_gather.log
ccl_mpi_hybrid_all_reduce.log
ccl_mpi_hybrid_send_recv.log

---

Checklist

Title, Branch, and Commits

• PR title follows Conventional Commits.
• Branch name follows <type>/xxx-yyyy-zzzz where <type> matches the PR title's Conventional Commits type and words are joined with hyphens.
• Each commit message follows Conventional Commits.
• This small PR is a single squashable commit.
• No stray merge commits are present; this PR is cleanly stacked on feat(nccl): add point-to-point communication #58.
• No fixup! / squash! / wip commits remain.

Scope and Design

• Changes are minimal and limited to NCCL AllGather and its reliable validation.
• No dead code, commented-out blocks, debug prints, or unowned TODOs were introduced.
• No unrelated formatting churn was introduced.
• The existing public C API signature is unchanged; this PR provides its NCCL implementation.

General Code Hygiene

• Comments are limited to non-obvious intent.
• Every modified and added file ends with a trailing newline.
• No trailing whitespace, inconsistent indentation, or mixed formatting styles remain.
• Identifiers in comments and error messages use Markdown syntax where applicable.
• Comments and error messages are in English.
• Comments and error messages follow repository sentence conventions.

C++ Specific

• Code follows the Google C++ Style Guide and the repository .clang-format.
• clang-format version 16.0.6 passed with --dry-run --Werror.
• No exceptions were introduced.
• Error and warning wording follows repository conventions.
• No constructor initializer-order issue was introduced.
• Required blank-line conventions are satisfied.

Python Specific

N/A. This PR does not change Python files.

Testing

• All applicable AllGather examples were built and run successfully on NVIDIA with NCCL and with the mixed OpenMPI/NCCL configuration described above.

Build, CI, and Tooling

N/A. This PR does not add a backend or device to auto-detection. The repository's C++ formatting gate was reproduced locally; hosted CI will run on the pushed branch.

Documentation

• The new validated usage is documented by examples/ccl/all_gather.cc and this PR description.
• No user-visible breaking change is introduced.

Security and Safety

• No secrets, access tokens, internal URLs, customer data, or personal hardware identifiers were committed.
• No third-party source code was added.
• Buffer capacities and the NCCL-defined in-place offset are validated by the two-GPU example; no uninitialized reads were introduced.