pub fn attn_mma_keys(hd: usize, vd: usize) -> usize
The keys a tile of the f16 tensor-core attention takes: 64, or 32 where heads or values pass 192 wide.