Rounding in '/blck_size' calculations

[AndrewGodfrey]

I was reading ggml_nbytes while researching #3644, and noticed:

        nbytes = tensor->ne[0]*tensor->nb[0]/blck_size;

This division rounds down.

I imagine this is usually irrelevant because ne[0] tends to be a multiple of blck_size. But it feels like it's just a matter of time before this causes problems, if it hasn't already.

For GGML_TYPE_Q8_0 (blck_size==32) maybe it hasn't. But in the code there's a definition of QK_K==256, and that doesn't divide evenly into 3200 or 8640.

A similar calculation is done in some other places too. e.g. ggml_new_tensor_impl has ne[0]/ggml_blck_size(type)

What I'd expect to see for a block calculation, would be numBlocks * blck_size, where numBlocks = (ne[0] + blck_size - 1) / blck_size.

P.S. In finetune.cpp there are plenty of tensors whose ne[0] isn't a multiple of 32. But, those tend to be f32 or f16 and so they don't go through this codepath.