[ROCm] tweak rocm wheel build workflow to avoid ci timeouts (#1091)

* try increasing number of workers to 12 to avoid ci timeouts

* try 16-core runner for rocm

* use 15 cores; build for all rocm versions

* revert many-rocm wheel build when getting CI signal; add comment about runner

.github/workflows/wheels.yml

@@ -60,7 +60,7 @@ jobs:
               if os == 'windows-8-core':
                 continue
               include.append(dict(
-                os=os,
+                os="16-core-ubuntu",  # use for ROCm wheels only to avoid CI timeouts
                 python=python,
                 torch_version=torch_version,
                 toolkit_type="rocm",

.github/workflows/wheels_build.yml

@@ -153,7 +153,7 @@ jobs:
           amdgpu-install -y --usecase=rocm --no-dkms
           echo "ROCM_PATH=/opt/rocm" >> ${GITHUB_ENV}
           echo "PATH=$PATH:/opt/rocm/bin" >> ${GITHUB_ENV}
-          echo "MAX_JOBS=7" >> ${GITHUB_ENV}
+          echo "MAX_JOBS=15" >> ${GITHUB_ENV}
 
       - name: Install dependencies
         run: $PY -m pip install wheel setuptools twine -r requirements.txt --extra-index-url $PYTORCH_INDEX_URL