ci: add concurrency cancellation to PR-triggered test workflows

[bigbitbus]

Why

CI minutes (Depot-hosted GitHub Actions runners, depot-ubuntu-*) have grown ~3–4× since Nov 2025. A primary structural cause: of all workflows, only codeflash.yml had a concurrency block. Every test workflow lacked cancel-in-progress, so pushing several commits to an open PR left all prior in-flight runs running and billed — we paid for one full matrix pass per intermediate commit instead of one per settled state.

What

Adds an identical workflow-level block to the 10 PR-triggered test workflows:

concurrency:
  group: ${{ github.workflow }}-${{ github.ref }}
  cancel-in-progress: ${{ github.event_name == 'pull_request' }}

• On PRs — a new push cancels the superseded run for that PR. You only pay for the latest commit's CI (the only result that gates merge).
• On push: [main] — the expression resolves to false, so runs queue instead of cancelling. Every merged commit keeps its own CI / required checks; nothing is orphaned.
• Group key is workflow + ref, so distinct PRs never cancel each other and different workflows on the same branch stay independent.

Files

unit_tests_inference_x86, unit_tests_workflows_x86, unit_tests_inference_sdk_x86, unit_tests_inference_cli_x86, unit_tests_inference_experimental, integration_tests_workflows_x86, integration_tests_inference_experimental_cpu, integration_e2e_tests_inference_sdk_x86, check_model_licenses, static_code_analysis.

codeflash already had a concurrency block and is unchanged.

Verification

• All 10 files parse as valid YAML.
• No duplicate workflow name: values exist, so no two workflows collide into one cancel-group.
• Purely additive — no job logic, runner class, or matrix touched. Cancelled runs report cancelled (not failed), so branch protection is unaffected.

Follow-ups (not in this PR)

• Add paths: filters so Python tests skip docs-only / unrelated-package changes.
• Right-size large Depot runners (integration_tests_workflows_x86 on -16, docs on -32).
• Gate docker.inference-exp full 6-image CUDA matrix builds on every push to main.

🤖 Generated with Claude Code