fix: support legacy CUDA graph instantiate

[voltjia]

Summary

• Keep the public GraphInstantiate(GraphExec*, Graph) runtime API while compiling against CUDA-compatible runtimes that expose either the newer two-argument cudaGraphInstantiate signature or the legacy five-argument signature.
• Apply the compatibility wrapper to src/native/cuda/nvidia/runtime_.h and src/native/cuda/iluvatar/runtime_.h.

Motivation

InfiniTensor/InfiniOps#790 fails in the Iluvatar unit CI after InfiniRT PR #9 because the job installs InfiniRT from master and builds with CoreX 4.3 / CUDA 10.2-compatible headers. In that environment, cudaGraphInstantiate is declared as:

cudaGraphInstantiate(cudaGraphExec_t*, cudaGraph_t, cudaGraphNode_t*, char*, size_t)

The generated InfiniRT dispatch validates and calls GraphInstantiate(GraphExec*, Graph), so the previous direct forwarding lambda fails to compile against those headers.

Related: #25 fixes the same downstream CI failure by removing unverified Iluvatar graph support. This PR is the alternative path if maintainers want to keep the Iluvatar graph API exposed and fix the compile-time signature mismatch instead.

Type of Change

• feat - new feature / new operator / new platform
• fix - bug fix
• perf - performance improvement (no behavioral change)
• refactor - code restructuring without behavior change
• test - adding or fixing tests only
• docs - documentation only
• build / ci - build system or CI configuration
• chore - tooling, formatting, or other non-code changes
• Breaking change (requires a ! in the Conventional Commits prefix or a BREAKING CHANGE: footer)

Platforms Affected

• CPU (WITH_CPU)
• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• PyTorch C++ bindings (WITH_TORCH)
• Build system / CMake / CI
• Python bindings / user-facing API

Smoke Test Result

git diff --check
# passed

C:\Users\huang\.cache\codex-runtimes\codex-primary-runtime\dependencies\python\python.exe -m py_compile scripts\generate_public_headers.py
# passed

C:\Users\huang\.cache\codex-runtimes\codex-primary-runtime\dependencies\python\python.exe scripts\generate_public_headers.py --output-dir C:\Users\huang\.codex\visualizations\2026\07\10\019f49e9-fdde-70f0-a3dc-1837693cfd9e\generated-check\include --source-output C:\Users\huang\.codex\visualizations\2026\07\10\019f49e9-fdde-70f0-a3dc-1837693cfd9e\generated-check\src\runtime_dispatch.cc --devices nvidia iluvatar
# passed

Test Results on Supported Platforms

Platform	Affected	Build / Smoke Result	Full Result / Notes
NVIDIA	Yes	Not run locally; PR CI clang-format, ruff, and docs build passed.	The wrapper keeps the two-argument CUDA path when that overload exists.
Iluvatar	Yes	Not run locally; this Windows workspace does not have cmake, a C++ compiler, CUDA, or the Iluvatar CoreX SDK in PATH.	Needs Iluvatar/CoreX CI validation if this approach is preferred over #25.
MetaX	No	N/A - not affected	N/A
Cambricon	No	N/A - not affected	N/A
Moore	No	N/A - not affected	N/A
Ascend	No	N/A - not affected	N/A

Full `pytest` output (optional)

N/A - backend pytest was not run locally.

Benchmark / Performance Impact

N/A - this is a compile-time compatibility fix for graph runtime API forwarding.

Notes for Reviewers

• fix: remove unverified Iluvatar graph support #25 is the conservative option if Iluvatar graph capture/replay should remain disabled until hardware behavior is validated.
• This PR preserves the Iluvatar graph API surface but only fixes the header signature mismatch; it does not prove Iluvatar graph replay correctness.
• If fix: remove unverified Iluvatar graph support #25 merges first, this PR should either be closed or reduced to any desired NVIDIA-only compatibility cleanup.

[voltjia]

Closing this for now because #25 provides the current conservative fix by removing the unverified Iluvatar graph runtime support. The compatibility-wrapper approach in this PR is not needed while #25 is the chosen path.