[fschlimb]

Build LLVM and tileir-to-mlir separately and re-use through cache.
Build and test the full python matrix.

[Copilot]

🔵 Needs a closer look

The LLVM/MLIR/LLD export and static-linkage changes need end-to-end native build/install and wheel-import results on regular and free-threaded Python, since the added regression test covers configuration only.

No open issues 🎉

What changed in this PR

Separates the CPU backend’s Python-independent native SDK from wheel builds, allowing CI to reuse LLVM and TileIRToMLIR across the expanded Python matrix.

Changes:

• Adds native SDK build, installation, and consumption paths.
• Caches the SDK and builds/tests wheels for Python 3.11–3.15, including free-threaded 3.14 and 3.15.
• Updates the TileIRToMLIR revision and adds a configuration test without Python or CUDA discovery.

File	Description
CMakeLists.txt	Adds a native-SDK-only build option.
backends/​test/​test_compiler_build.py	Tests SDK configuration without Python or CUDA discovery.
backends/​cpu/​compiler.cmake	Uses installed LLD headers when consuming the SDK.
backends/​compiler.cmake	Adapts compiler includes, definitions, output, and installation for SDK builds.
backends/​CMakeLists.txt	Supports building or consuming the native SDK and installing licenses.
backends/​cmake/​TileIRToMLIR.cmake	Configures Python-independent SDK distribution builds.
backends/​cmake/​FetchTileIRToMLIR.cmake	Updates the pinned TileIRToMLIR revision and checksum.
.github/​workflows/​ci-cpu.yml	Adds SDK caching and expands wheel builds and tests across Python versions.

_{Review effort: Balanced}