tree-sitter-languages: add build-tree-sitter-languages.yml for riscv64 wheels

[luhenry]

Adds .github/workflows/build-tree-sitter-languages.yml, building tree-sitter-languages 1.10.2 riscv64 wheels for cp312/cp313/cp314/cp314t. Upstream publishes real per-interpreter compiled wheels (cp37–cp312, ~8.5 MB each) for x86_64/i686/aarch64/musl/macOS/Windows but nothing for riscv64.

The workflow mirrors upstream's release.yml: cibuildwheel over a checkout of the tag, with before-build running build.py, which shallow-clones the 48 grammar repositories pinned in repos.txt and links their parser.c/scanner.cc into a single tree_sitter_languages/languages.so (~88 MB unstripped) that the Cython core extension dlopens through py-tree-sitter. Tests are upstream's own pytest --showlocals {package}/tests, which parses Python with the built parser and instantiates all 48 grammars.

Upstream pins only Cython, and three build inputs had to be pinned to build the tag at all today — all three failures are arch-independent and reproduce on x86/arm:

• tree-sitter<0.22 — build.py calls Language.build_library and core.pyx calls Language(path, name); both were removed in tree-sitter 0.22, so an unpinned resolve (upstream's pip install -e .) gets 0.26 and neither the build nor the test suite can work. The pin covers the test phase too.
• cython==3.3.0 — upstream's cython==3.0.8 predates free threading and fails to compile core.pyx under cp314t; 3.3.0 builds cleanly on all four interpreters (core.pyx is 15 lines of plain Python, so there is no codegen-drift exposure of the kind gotcha 23 describes).
• setuptools<79 — setuptools 79.0.0 regressed distutils' C++ link command: the class default for linker_exe_cxx is ["c++", "-shared"], and _linker_params strips that whole prefix off linker_so_cxx, so -shared is dropped and build.py's link of languages.so dies with crt1.o: undefined reference to 'main'. build.py uses new_compiler() without customize_compiler(), so LDCXXSHARED never gets a look-in and an env var cannot work around it. Bisected in the manylinux image: 78.1.1 emits c++ -shared …, 79.0.1 emits c++ ….

patches/tree-sitter-languages/1.10.2/0001-ship-the-vendored-grammar-licences-in-the-wheel.patch closes a licensing gap: languages.so is 48 third-party grammars compiled in, each under its own MIT/Apache-2.0 terms, while the wheel ships only tree-sitter-languages' own LICENSE. The patch adds license_files=['LICENSE', 'vendor/*/LICENSE*'] to setup.py (the default glob only looks at the project root), which lands 46 licence files in dist-info/licenses/vendor/ — the two exceptions, tree-sitter-ql and tree-sitter-fixed-form-fortran, carry no licence file at the commits repos.txt pins. The test command asserts the count so the patch cannot silently stop applying.

Local validation before pushing: a full cibuildwheel --only cp312-manylinux_aarch64 run against quay.io/pypa/manylinux_2_39_aarch64 (same Rocky 10 image family) — clone, compile, wheel, auditwheel repair to manylinux_2_34_aarch64, licence assertion, 3 tests passed — plus the same build-and-test loop by hand in that image for cp313, cp314 and cp314t. cp314t works but re-enables the GIL on import (Cython does not declare Py_mod_gil for this module), which is a runtime property of upstream's source, not of the port.