One activation row; gate and up at the same or different widths, decoding through a threadgroup codebook
(kernels/trellis/gate-up-codebook.ts) with fusedGateUpSwiglu's (same widths) and fusedGateUpSwigluMixed's (different
widths) output bits. mixedTail is the SwiGLU arithmetic where the widths differ (fusedGateUpSwigluMixed's tail);
same widths take the float32 sigmoid.
One activation row; gate and up at the same or different widths, decoding through a threadgroup codebook (kernels/trellis/gate-up-codebook.ts) with fusedGateUpSwiglu's (same widths) and fusedGateUpSwigluMixed's (different widths) output bits.
mixedTailis the SwiGLU arithmetic where the widths differ (fusedGateUpSwigluMixed'stail); same widths take the float32 sigmoid.