Strix Halo: one job per model via 3 slices, RUN_SLOW=1, default timeout + pytest-timeout
Replicate MI300/MI355: split tests/models into 3 slices (~172 each, under
the 256-job matrix cap), one models-ci job group per slice -> one job per
model in a single run (job name = model). Enable the full slow suite
(RUN_SLOW=1). Remove the timeout-minutes:240 override (GitHub 6h default)
and add pytest-timeout (--timeout=1200) to kill hung tests early.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>