Hold the Ethos-U backend suites at 8 pytest workers

.github/workflows/_test_backend.yml

@@ -46,6 +46,14 @@ on:
         required: false
         type: string
         default: ci-image:executorch-ubuntu-22.04-clang12
+      pytest-workers:
+        description: |
+          pytest-xdist worker count. Defaults to the pod's CPU allowance, which
+          is right unless a suite needs more memory per worker than the label's
+          8GiB per core.
+        required: false
+        type: string
+        default: ''
 
 jobs:
   docker-image:
@@ -77,6 +85,10 @@ jobs:
       script: |
         set -eux
 
+        if [[ -n "${{ inputs.pytest-workers }}" ]]; then
+          export PYTEST_XDIST_AUTO_NUM_WORKERS="${{ inputs.pytest-workers }}"
+        fi
+
         source .ci/scripts/test_backend.sh "${{ matrix.suite }}" "${{ matrix.flow }}" "${RUNNER_ARTIFACT_DIR}"
 
   package-golden-artifacts:

.github/workflows/test-backend-arm.yml

@@ -35,6 +35,12 @@ jobs:
       timeout: 120
       run-linux: true
       docker-image: ci-image:executorch-ubuntu-22.04-arm-sdk
+      # The Ethos-U flows run a Corstone FVP per worker and the failures each
+      # emit a Vela dump of a few hundred thousand lines, which killed the pod
+      # once `auto` stopped discounting hyperthreads: linux.4xlarge.memory gave
+      # 8 workers 16GiB each, the OSDC label gives 15 workers 7.7GiB each. 8
+      # keeps what the suite had.
+      pytest-workers: 8
 
   test-arm-vgf:
     uses: ./.github/workflows/_test_backend.yml