feat(nccl): add point-to-point communication

[voltjia]

Summary

Add NVIDIA NCCL implementations for infinicclSend and infinicclRecv, while preserving communicator-aware fallback to an available backend. The branch is stacked on #57 so provider header changes reliably regenerate the generated bridge.

Changes

• NCCL point-to-point support
  • Add common CCL validation and NCCL Send/Recv providers.
  • Forward dtype, peer rank, communicator, and CUDA stream to ncclSend and ncclRecv.
• Safe communicator dispatch
  • Select only backends and devices supported by the requested operation.
  • Fall back from an unsupported communicator backend to another active communicator backend.
  • Return kNotSupported instead of entering an unavailable template dispatch and aborting.
• Examples and tests
  • Add a deterministic dual-GPU send/receive example.
  • Add backend-selection tests for supported, fallback, and unsupported communicator combinations.

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. No performance claim is made; the change adds a previously unavailable NCCL operation.

Known Issues & Future Work

• This stacked PR depends on fix(build): regenerate communication bridge for header changes #57. Rebase the branch onto master after fix(build): regenerate communication bridge for header changes #57 merges.
• This PR adds NCCL point-to-point communication only; it does not add other missing NCCL collectives.

Test Results

Test environment: ssh nvidia, accelerator-dev/nvidia:latest, two NVIDIA A100-SXM4-80GB GPUs.

• NVIDIA+NCCL Release configure and full example build passed.
• CTest passed, including supported, fallback, and unsupported backend-selection cases.
• examples/ccl/send_recv passed on two A100 GPUs and received the expected value 7.0.
• Existing examples/ccl/all_reduce passed as a regression check.
• OpenMPI+NCCL Release configure and full example build passed.
• mpirun -np 2 examples/mpi/send_recv passed and received the expected data.
• clang-format 16.0.6 and git diff --check passed.

Test Involved Platform

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

Test Involved Backend

• OpenMPI
• MPICH
• NCCL
• MCCL

---

Checklist

Title, Branch, and Commits

• PR title follows Conventional Commits.
• Branch name follows <type>/xxx-yyyy-zzzz.
• Each commit message follows Conventional Commits.
• Relative to the stacked base, this small PR is a single squashable commit.
• No stray merge commits from master.
• No fixup! / squash! / wip commits remain.

Scope and Design

• Changes are minimal and contain no unrelated modifications.
• No dead code, debug output, or unowned TODOs were added.
• No unrelated formatting churn was introduced.
• The public Send/Recv behavior is intentional and covered by examples and tests.

General Code Hygiene

• Comments are limited to non-obvious intent.
• Every modified or added file ends with a single trailing newline.
• No trailing whitespace, inconsistent indentation, or mixed formatting remains.
• Code identifiers in comments and errors use Markdown backticks where applicable.
• Comments and error messages are in English.
• Comments and error messages follow the repository language conventions.

C++ Specific

• Code follows the Google C++ Style Guide.
• clang-format 16.0.6 passes on all modified C++ files.
• No exceptions are thrown.
• Error handling follows repository conventions.
• Constructor initializer order is unchanged or matches declarations.
• Class and function spacing follows repository conventions.
• Class member spacing follows repository conventions.
• Namespace spacing follows repository conventions.

Python Specific

N/A: No Python files are changed by this PR.

Testing

• Applicable NCCL and OpenMPI examples were built and tested successfully in the environments listed above.

Build, CI, and Tooling

• N/A: No new backend or device auto-detection entry is required.
• The applicable clang-format check passes; Ruff is not applicable.

Documentation

• The new example documents the public point-to-point usage.
• N/A: No user-visible breaking change is introduced.

Security and Safety

• No secrets, internal URLs, customer data, or personal hardware identifiers are included.
• No third-party code is introduced.
• Provider inputs, communicator backend, device, and operation support are validated before dispatch.