ci: add ci_test workflows and fix NVIDIA probing

[zhangyue207]

Summary

• Move the ci_test CI work into master.
• Replace the in-tree .ci implementation with the external InfiniTensor/ci submodule.
• Add .github/workflows/ci_test.yml and .github/ci_config.yaml as the parent GitHub Actions entrypoint/config.
• Run the NVIDIA job with --devices nvidia.
• Make tests/conftest.py respect explicit --devices platform selections before falling back to the historical broad torch-device mapping.

Motivation

This PR brings the ci_test workflow work into master and fixes the NVIDIA validation failure found while testing that workflow locally.

The NVIDIA job crashed because device=cuda was treated as all CUDA-like InfiniOps platforms: nvidia, metax, and iluvatar. On an NVIDIA runner, skip_op_without_platform_impl could still query MetaX / Iluvatar implementation indices, which may abort in native code instead of producing a safe pytest skip.

Closes: N/A.

Type of Change

• fix — bug fix
• build / ci — build system or CI configuration

Platforms Affected

• NVIDIA (WITH_NVIDIA)
• Iluvatar (WITH_ILUVATAR)
• MetaX (WITH_METAX)
• Cambricon (WITH_CAMBRICON)
• Moore (WITH_MOORE)
• Ascend (WITH_ASCEND)
• Build system / CMake / CI

Test Results on Supported Platforms

Platform	Built	pytest Result	Notes / Hardware
NVIDIA	Yes	3758 passed, 1363 skipped	Local self-hosted NVIDIA flow via python3 .ci/run.py --config .github/ci_config.yaml --job nvidia_gpu --gpu-id 0 --local.
Iluvatar	No	Not run	No Iluvatar hardware available on this machine.
MetaX	No	Not run	No MetaX hardware available on this machine.
Cambricon	No	Not run	No Cambricon hardware available on this machine.
Moore	No	Not run	No Moore hardware available on this machine.
Ascend	No	Not run	No Ascend hardware available on this machine.

Benchmark / Performance Impact

N/A — this PR only changes CI/test platform selection behavior.

Notes for Reviewers

This PR is now targeted at master, so it includes the original ci_test branch changes plus the follow-up NVIDIA probing fix.

Review focus:

• The .ci submodule migration and .github workflow/config boundary.
• With explicit --devices nvidia, device=cuda only queries nvidia implementation indices.
• Without explicit --devices, the historical broad fallback remains unchanged.

---

Checklist

• PR title and commit messages follow Conventional Commits.
• Branch name follows <type>/xxx-yyyy-zzzz. Existing branch is codex/fix-nvidia-platform-probing.
• Commits are meaningful and no fixup! / squash! / wip commits remain.
• Changes are scoped to moving ci_test CI into master and fixing the NVIDIA CI failure.
• ruff check tests/conftest.py passes.
• ruff format --check tests/conftest.py passes.
• python3 -m py_compile tests/conftest.py passes.
• Matrix generation was checked with python3 .ci/config_to_matrix.py --config .github/ci_config.yaml --dump-by-type.
• pytest was run locally on every supported platform this PR can affect. NVIDIA was run locally; other hardware was not available.
• Standalone regression test added. No standalone regression test was added; validation is through the NVIDIA CI job path that previously crashed.
• No secrets, third-party code, or unsafe native-code changes were added.