update llama

[ftynse]

• fix rmsnorm to compute what it should
• use complex numbers instead of rotation matrix
• don't append

[ftynse]

Adapting this from another version that is known to run on proper data. Note that the weights for dense layers coming from pytorch are transposed and my version is using x @ weights.mT whereas this version uses weights @ x.