feat: add point-to-point communication for mccl

[lqinfdim]

Summary

Add point-to-point communication support for the MCCL backend.

This PR introduces Send and Recv operations for MCCL on MetaX GPUs,
including the backend adapters, common CCL implementations, bridge
registration, tests, and a runnable example.

Changes

• MCCL backend

  • Add MCCL adapters for mcclSend and mcclRecv in
    src/backends/ccl/mccl/impl/send.h and
    src/backends/ccl/mccl/impl/recv.h.
  • Add MCCL API declarations for point-to-point communication in
    src/backends/ccl/mccl/api.h.
  • Register the MCCL Send and Recv implementations with the CCL bridge.

• Common CCL implementation

  • Add common point-to-point operation implementations in
    src/backends/ccl/common/impl/send.h and
    src/backends/ccl/common/impl/recv.h.
  • Preserve the existing backend and device validation behavior.

• Build, tests, and example

  • Update bridge generation dependencies for the new point-to-point
    operations.
  • Add point-to-point implementation tests in
    tests/ccl_point_to_point_impl.cc.
  • Add bridge dependency and bridge generation coverage.
  • Add the examples/ccl/send_recv.cc example.
  • Validate command-line arguments without exception-based parsing.

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

This PR adds point-to-point communication support and does not change the
performance characteristics of existing collective operations.

Known Issues & Future Work

• The implementation currently covers MCCL point-to-point send and receive
  operations.
• Multi-device MCCL runtime validation depends on a working MetaX/MCCL
  runtime environment and communicator initialization.
• Additional benchmark coverage can be added in a follow-up change if
  point-to-point performance tracking is required.

Test Results

The following checks were run with MetaX and MCCL enabled:

• Clean CMake build completed successfully.
• CTest passed: 3/3 tests.
  • bridge_dependency_contract
  • bridge_generation
  • ccl_point_to_point_impl
• The examples/ccl/send_recv executable was built successfully.
• Invalid command-line arguments return an error without throwing exceptions.
• Help output is returned successfully with -h.

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

The following items should be checked after verifying them locally.

Title, Branch, and Commits

• PR title follows Conventional Commits.
• Branch name follows the repository naming convention:
  feat/mccl-point-to-point-20260818.
• Commit message follows Conventional Commits.
• This small PR is maintained as a single squashable commit.
• No stray merge commits from master.
• No fixup!, squash!, or wip commits remain.

Scope and Design

• Changes are limited to MCCL point-to-point communication and its
  tests/examples.
• No unrelated backend or platform behavior was added.
• No dead code, debug output, or unexplained TODOs were introduced.
• No unrelated formatting churn was intentionally introduced.

General Code Hygiene

• Modified files use English comments and error messages.
• Modified files have trailing newlines.
• git diff --check passes.

C++ Specific

• No exceptions are thrown by the new example argument parsing or
  point-to-point implementation.
• clang-format version 16 has been run on all modified C++ files.
  
• Existing constructor and validation conventions are preserved.

Testing

• The relevant CCL example was built successfully.
• The relevant unit and bridge tests passed.

Build, CI, and Tooling

• CI workflows are green.
  

Documentation

• Documentation update is not required for this internal backend
  implementation and example addition.

Security and Safety

• No secrets, tokens, internal URLs, customer data, or hardware
  identifiers were committed.
• No unsafe pointer arithmetic or uninitialized reads were introduced.