[ci] Adjust expect_failure and expect_pytorch_failure logic

[ScottTodd]

Motivation

Handling a few related issues that will complicate multi-arch release pipeline work (#3334):

1. expect_failure plumbing

expect_failure was optionally set on a build variant (e.g. release/asan/tsan). As of 3c528c0, it is never set now.

We were passing expect_failure all the way through to multi_arch_build_portable_linux.yml where stages are built, but it was ignored there. That workflow doesn't have a single job that it could mark as continue-on-error as the job that it replaces:

TheRock/.github/workflows/build_portable_linux_artifacts.yml

Lines 61 to 67 in 740c338

	jobs:
	build_portable_linux_artifacts:
	name: Build (xfail ${{ inputs.expect_failure }})
	# azure-linux-scale-rocm are used for regular CI builds
	# azure-linux-scale-rocm-heavy are used for CI builds that require more resources (ex: ASAN and TSAN builds)
	runs-on: ${{ (contains(inputs.build_variant_label, 'asan') || contains(inputs.build_variant_label, 'tsan')) && 'azure-linux-scale-rocm-heavy-ramdisk' || 'azure-linux-scale-rocm' }}
	continue-on-error: ${{ inputs.expect_failure }}

Solution: stop handling expect_failure altogether. We can add it back later as needed (somehow).

2. expect_pytorch_failure plumbing

expect_pytorch_failure was being read from the build variant when it actually exists on a GPU family:

TheRock/build_tools/github_actions/amdgpu_family_matrix.py

Lines 178 to 184 in 96c9642

	"windows": {
	"test-runs-on": "",
	"family": "gfx900",
	"fetch-gfx-targets": [],
	"build_variants": ["release"],
	"expect_pytorch_failure": True,
	},

TheRock/build_tools/github_actions/configure_multi_arch_ci.py

Line 861 in 96c9642

expect_pytorch_failure = variant_config.get("expect_pytorch_failure", False)

In multi-arch CI we currently build for all GPU families in a single workflow, with a matrix in build_pytorch_wheels_per_family. We could do two layers of filtering here: one for "should this workflow run build pytorch at all?" and another for "does the pytorch build for this GPU family work?"

Solution: correct the configuration plumbing, add TODO to handle in the matrix somehow

Technical Details

I included a few other loosely related / incidental changes here (could move to a separate PR on request):

• Reordered arguments to expand_build_configs
• Folded TestDualLabelRunnerSelection tests into TestExpandBuildConfigs where they belong (and can reuse helper functions)

Test Plan

• Updated unit tests
• Watch CI runs, including with a target like gfx906 that fails to build pytorch (would fail before, should still fail)

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[ScottTodd]

Test Plan

• Watch CI runs, including with a target like gfx906 that fails to build pytorch (would fail before, should still fail)

Uh, it succeeded: https://github.com/ROCm/TheRock/actions/runs/24469430596/job/71558513727?pr=4500 . Maybe #1927 is fixed now?

TheRock/build_tools/github_actions/amdgpu_family_matrix.py

Lines 197 to 204 in 1db2c09

	# TODO(#1927): Resolve error generating file `torch_hip_generated_int4mm.hip.obj`, to enable PyTorch builds
	"windows": {
	"test-runs-on": "",
	"family": "gfx906",
	"fetch-gfx-targets": [],
	"build_variants": ["release"],
	"expect_pytorch_failure": True,
	},

This PR also has merge conflicts now. I'll probably need to split it into smaller PRs.