Update Sonoma model with faster 8x8 conv and split einsum attention

Update sequoia mode with transposed value cache and 4:508 input:cache length

Upload Sequoia model

Update README.md

Add model

initial commit