ci: use ARM runner for Python ARM release builds (#3411)

We can get both 2_17 and 2_28 working if we use an ARM runner.

Author: wjones127

.github/workflows/pypi-publish.yml

@@ -15,18 +15,20 @@ jobs:
           - platform: x86_64
             manylinux: "2_17"
             extra_args: ""
+            runner: ubuntu-22.04
           - platform: x86_64
             manylinux: "2_28"
             extra_args: "--features fp16kernels"
+            runner: ubuntu-22.04
           - platform: aarch64
             manylinux: "2_17"
             extra_args: ""
+            runner: ubuntu-2404-4x-arm64
           - platform: aarch64
             manylinux: "2_28"
             extra_args: "--features fp16kernels"
-          # We don't build fp16 kernels for aarch64, because it uses
-          # cross compilation image, which doesn't have a new enough compiler.
-    runs-on: "ubuntu-22.04"
+            runner: ubuntu-2404-4x-arm64
+    runs-on: ${{ matrix.config.runner }}
     steps:
     - uses: actions/checkout@v4
       with: