Retry download and install NDK when testing Android (pytorch#97067)

As this step uses network to download and install NDK, it could fail flakily, i.e. https://github.com/pytorch/pytorch/actions/runs/4452757793/jobs/7820701670.  So I'm adding retrying to the workflow.

I could try figure out a way to Dockerize this, but not sure yet how to handle the GitHub action `reactivecircus/android-emulator-runner@v2` in Docker.  So let's opt for the easy fix of retrying.
Pull Request resolved: pytorch#97067
Approved by: https://github.com/malfet

.github/workflows/_run_android_tests.yml

@@ -53,18 +53,39 @@ jobs:
           python-version: 3.8
           environment-file: .github/requirements/conda-env-${{ runner.os }}-${{ runner.arch }}
 
-      - name: Build PyTorch Android
-        run: |
-          # Install NDK 21 after GitHub update
-          # https://github.com/actions/virtual-environments/issues/5595
-          ANDROID_ROOT="/usr/local/lib/android"
-          ANDROID_SDK_ROOT="${ANDROID_ROOT}/sdk"
-          SDKMANAGER="${ANDROID_SDK_ROOT}/cmdline-tools/latest/bin/sdkmanager"
-          echo "y" | ${SDKMANAGER} "ndk;21.4.7075529"
+      - name: Install NDK
+        uses: nick-fields/[email protected]
+        with:
+          timeout_minutes: 5
+          max_attempts: 3
+          retry_wait_seconds: 90
+          command: |
+            # Install NDK 21 after GitHub update
+            # https://github.com/actions/virtual-environments/issues/5595
+            ANDROID_ROOT="/usr/local/lib/android"
+            ANDROID_SDK_ROOT="${ANDROID_ROOT}/sdk"
+            ANDROID_NDK="${ANDROID_SDK_ROOT}/ndk-bundle"
+            ANDROID_NDK_VERSION="21.4.7075529"
+
+            SDKMANAGER="${ANDROID_SDK_ROOT}/cmdline-tools/latest/bin/sdkmanager"
+            # NB: This step downloads and installs NDK, thus it could be flaky.
+            # However, SDKMANAGER doesn't return a non-zero status code when it
+            # happens despite the fact that the corrupted file that it has isn't
+            # a ZIP archive and couldn't be extracted
+            echo "y" | ${SDKMANAGER} "ndk;${ANDROID_NDK_VERSION}"
 
-          export ANDROID_NDK="${ANDROID_SDK_ROOT}/ndk-bundle"
-          ln -sfn ${ANDROID_SDK_ROOT}/ndk/21.4.7075529 ${ANDROID_NDK}
+            ln -sfn "${ANDROID_SDK_ROOT}/ndk/${ANDROID_NDK_VERSION}" "${ANDROID_NDK}"
+            # So, we need to manually verify the existence of NDK afterward
+            # and return a failure if the file isn't there
+            if [ ! -f "${ANDROID_NDK}/build/cmake/android.toolchain.cmake" ]; then
+              exit 1
+            fi
 
+            echo "ANDROID_SDK_ROOT=${ANDROID_SDK_ROOT}" >> "${GITHUB_ENV}"
+            echo "ANDROID_NDK=${ANDROID_NDK}" >> "${GITHUB_ENV}"
+
+      - name: Build PyTorch Android
+        run: |
           echo "CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname "$(which conda)")/../"}" >> "${GITHUB_ENV}"
           ${CONDA_RUN} ./scripts/build_pytorch_android.sh x86