clip: fix nb calculation for qwen3-vl

[ngxson]

This may silently fail if mul_mat return something other than F32 in the future, so I'm using ggml_row_size to be more future-proof.

[CISC]

Did I miss something? I don't think we'll ever expect it to return quantized tensors, so this is surely overkill? If expecting anything other than F32 (in which case, we need to fix up a lot of places) using d_head*ggml_element_size(cur) should be enough?

[ngxson]

@CISC honestly I don't know if hard-coding sizeof(float) is a good practice in GGML overall, even when the tensor is guaranteed to be F32 (who knows, maybe a novel backend will store the tensor using just 31 bits instead of full 32 bits? hmm well bad example, in this case other things will break before this 😄 )

ggml_element_size(cur) * N is also OK, indeed I don't know what's the recommendation.

Not sure if you already came across this pattern somewhere else @ggerganov , maybe in whisper.cpp or somewhere else. What's your recommendation here?