workflows: add build-apache-tvm-ffi.yml for riscv64 manywheel builds

[luhenry]

What

Adds .github/workflows/build-apache-tvm-ffi.yml to build and publish riscv64 wheels for apache-tvm-ffi (Apache TVM FFI — an open ABI/FFI for ML systems) to pypi.riseproject.dev.

Approach

sdist → bdist (abi3). The PyPI sdist is self-contained: it bundles the dlpack and libbacktrace submodules, the test suite, and a setuptools-scm-derived _version.py. So the riscv64 wheel builds from it with no git history, submodules, or codegen toolchain — just scikit-build-core + Cython + CMake + a C++17 compiler, all present in the manylinux riscv64 image.

• The sdist is arch-independent, so it's built once on ubuntu-latest; only the bdist runs on ubuntu-24.04-riscv.
• Wheels are abi3 (pyproject sets wheel.py-api = "cp312"), so the matrix is cp312 (loadable on 3.12/3.13/3.14) + a free-threaded cp314t build — the two entries of upstream's cibuildwheel build list that actually differ.
• CIBW_ARCHS: riscv64 overrides pyproject's [tool.cibuildwheel.linux] archs = ["x86_64", "aarch64"].

Testing (mirrors upstream)

The build frontend (build[uv]), test-command (pytest {package}/tests/python) and test-groups (["test"] → torch/numpy/ml_dtypes/pytest/ninja) are all inherited from the sdist's pyproject.toml, so wheels are tested exactly the way upstream tests them. Those test deps have no riscv64 wheels on public PyPI, so pip/uv are pointed at our registry. only-binary is scoped to the test phase only (via CIBW_TEST_ENVIRONMENT) — the build backend needs cython>=3.0, which has no riscv64 wheel and must compile from its sdist, so a global only-binary would break the build. GPU-only tests self-skip via torch.cuda.is_available().

Local validation (manylinux riscv64 container, under emulation)

• ✅ The self-contained sdist builds apache_tvm_ffi-0.1.12-cp312-abi3-linux_riscv64.whl.
• ✅ Test-group deps install wheels-only from the registry: torch 2.13.0+cpu, numpy, ml_dtypes, pytest.
• ✅ actionlint clean except the expected unknown-label warning for the self-hosted ubuntu-24.04-riscv runner.

🤖 Generated with Claude Code

[luhenry]

@justeph @threexc could I please get a review? Thanks!

[threexc]

Provided the current builds complete, LGTM. We'll want to comment about this in the docs PR though:

      # Mirror upstream's test-command (pytest {package}/tests/python -vvs)
      # but drop two tests that fail only because of this riscv64 runner's
      # environment, not the wheel (2300+ others pass; cp314t, which pulls no
      # torch, is fully green):
      #   * test_libinfo.py - every test there shells out to `tvm-ffi-config`
      #     and asserts the subprocess's stderr is empty. Importing tvm_ffi
      #     pulls in torch, whose bundled libcpuinfo can't parse this runner's
      #     /sys CPU topology (core_id reads "-1") and prints
      #     "Error in cpuinfo: failed to parse ... core_id" to stderr. The
      #     noise is nondeterministic across the module's subprocess tests, so
      #     the whole file is dropped via --ignore (an absolute path, so it's
      #     unambiguous regardless of pytest's rootdir).
      #   * test_filelock.py::test_concurrent_access - spawns 16 concurrent
      #     torch-importing workers under a hard 60s wait; the slower riscv64
      #     runner blows the timeout. Dropped with -k by its (unique) name -
      #     rootdir-independent, unlike a path-based --deselect (which silently
      #     no-ops when the nodeid prefix doesn't match pytest's rootdir).