pub const FP8_GROUPED_TC_GATE_UP_COLS_PER_CTA: u32 = 64;
2026-09-28: Output columns per tensor-core gate+up CTA (TC_GU_COLS).
TC_GU_COLS