ggml : storing strides as number of elements instead of number of bytes

[slaren]

Currently, we store the strides between elements of each dimension as a number of bytes in ggml_tensor::nb. In practice, this complicates code because strides always need to be multiplied by the element size, and accessing elements requires first casting the pointers to char *.

I am not sure if there are any cases where we would want a byte stride that isn't a multiple of the element size, as this would mean that the addresses would no longer be aligned to the element size, which is not ok in many platforms. Therefore I think we could simplify the code a bit by storing strides as numbers of elements instead of numbers of bytes.

[ggerganov]

Yup, I guess it would be an improvement. Probably we can rename nb to ns so that we get errors in all places in the codebase when refactoring this. And leave a comment explaining to 3rd party devs how to update if they are using nb somewhere in their projects