Strix Halo: throttle models matrix to max-parallel 7
The 256-job all-models run flooded the pool's queue: ~78 jobs ran, then
178 sat queued with 0 runners and the run stalled. The pool serves ~7
concurrently, so cap dispatch at max-parallel: 7 -- GitHub keeps the rest
pending instead of dumping them all into the runner queue at once.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>