fix: use correct BLOCK_SIZE in matmul function for ggml type

[xffxff]

To compute k_in_lhs_blocks, we should use T::VecDottype::BLOCK_SIZE because we obtain lhs_b using T::VecDotType::from_float. Similarly, to calculate k_in_rhs_blocks, we need to use T::BLOCK_SIZE since rhs_t is a slice of T (I'm just starting to learn quantization, so I'm not sure if it's correct, but it seems intuitive to me)