Add a test matrix job that runs on self-hosted GPU runners (AWS EC2
Ampere instances). Tests run inside Docker containers with --gpus all
using the pre-built test images from GHCR. Also update all image tags
to 2026-03-18 builds which include tileiras 13.2 (adds sm_86 support).

Docker containers run as root, so files created during tests (e.g.
.pytest_cache) are root-owned. Subsequent jobs on the same runner
fail when actions/checkout tries to clean the workspace. Fix by
restoring ownership after each test run.

Replace manual docker run/mount/permission-fix workflow with GitHub
Actions' built-in container: directive and --gpus all via options.

- actions/upload-artifact v4 -> v6
- actions/download-artifact v4 -> v7
- dorny/test-reporter v2 -> v3

- Add workflow-level permissions (contents: read, packages: read)
- Add checks: write to test job for dorny/test-reporter
- Add fail-fast: false to build and test matrices
- Replace if: always() with if: !cancelled() on test result steps

- Update test image tags to include git (needed by dorny/test-reporter)
- Remove --ignore internal (no internal folder in OSS)
- Remove -m 'not benchmark and not use_mlir' (run all tests, use_mlir
  auto-skips)

- Exclude benchmarks from the regular test job with -m "not benchmark"
  to prevent GPU OOM from large tensor allocations competing with
  parallel test jobs
- Add a dedicated benchmark job (Python 3.10 only, continue-on-error)
  mirroring the GitLab CI pattern
- Add git safe.directory config to the test job to fix dorny/test-reporter
  failing with git exit code 128 in Docker containers

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

Benchmark was running in parallel with all 4 test jobs, putting 5
concurrent GPU workloads on the same runner and causing OOM for
everything. Sequencing benchmark after test ensures it gets the
GPU to itself.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

The dorny/test-reporter steps were failing the job a second time when
tests failed. The pytest step is the authoritative failure signal;
the reporter is for display only.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>