Support broadcast addition for fp32

[li-plus]

This PR supported broadcast addition for two fp32 tensors as long as they are broadcast-able in row dimension. No longer requires same shape. Implemented on both CPU and CUDA.

[slaren]

For CUDA, it may be better to do this with a single kernel launch, similar to the way broadcasting is implemented for mul. @JohannesGaessler what do you think?

[JohannesGaessler]

Each CUDA kernel launch adds overhead. It is therefore preferable to launch as few kernels as reasonably possible. So I agree with slaren that an implementation that does the broadcasting with a single kernel launch would be preferable.

[li-plus]

Thanks for your advice! You are correct. Now I use a single kernel to compute broadcast add. This leads to huge performance improvement in linear layers.

[ggerganov]

Nice change!

Would be great if we utilize that ggml_add() is broadcast-able and simplify the inference of some of the example models. For example, no longer need to use ggml_repeat():

https://github.com/ggerganov/whisper.cpp/blob/4774d2feb01a772a15de81ffc34b34a1f294f020/examples/talk/gpt-2.cpp#L469-L471

https://github.com/ggerganov/whisper.cpp/blob/4774d2feb01a772a15de81ffc34b34a1f294f020/examples/talk/gpt-2.cpp#L449-L453

Also, need to handle the Metal and OpenCL implementations - add GGML_ASSERT(false && "not implemented") for now