Pallas kernels for Linear Attention, Mamba-2 and GLA, and the short causal convolution before them, on TPU v6e.