The compressed state of one latent attention layer after an append, as
caller-owned views of every stored position in order (float32). The caller
disposes all three. Rows follow the cache's row layout: right-justified, row
b's positions starting at leftPad[b].
The compressed state of one latent attention layer after an append, as caller-owned views of every stored position in order (float32). The caller disposes all three. Rows follow the cache's row layout: right-justified, row
b's positions starting atleftPad[b].