_MATRIX_JOB_RE/_variants_from_jobs grouped jobs into matrix cells by (py, ref) only, with no architecture dimension

[cgoea]

When a single PyTorch/JAX build cell fanned out test jobs across multiple GPU architectures (e.g. gfx942 and gfx1101), _refresh_same_run_fanout_tests broadcast one shared, worst-of variant list onto every architecture's test leaf — so a gfx1101 failure would flip gfx942's otherwise-passing leaf to failure too.