ci: cache less of llvm (#102)

Changes:
- Only cache the build artifacts of LLVM that we actually use to build
the wheels
- This means a ~9x reduction in cache usage for windows and ~3x for
linux

<img width="1067" height="317" alt="Screenshot 2025-12-11 at 09 34 31"
src="https://github.com/user-attachments/assets/dbafabff-9c28-4fa6-a23e-5a287590f91d"
/>


Link to PR:
Arm-Debug/vgf-adapter-model-explorer#102

Author: awphi

.github/actions/build-wheel/action.yml

@@ -120,7 +120,11 @@ runs:
       uses: actions/cache@v4
       id: cache-llvm
       with:
-        path: c/dependencies/llvm-project/b
+        path: |
+          c/dependencies/llvm-project/b/tools/mlir/python_packages
+          c/dependencies/llvm-project/b/tools/mlir/python
+          c/dependencies/llvm-project/b/bin
+
         key: llvm-${{ steps.dep-revisions.outputs.llvm-ref }}-${{ runner.os }}-${{ runner.arch }}-${{ inputs.python-version }}
 
     - name: Build LLVM