ggml : deprecate `GGML_TASK_FINALIZE`

[ggerganov]

The initial design of the compute tasks in ggml was each one to have 3 separate stages:

• GGML_TASK_INIT
• GGML_TASK_COMPUTE
• GGML_TASK_FINALIZE

So far, the GGML_TASK_FINALIZE step has been left completely unused and it seems it won't find any applications in the future. Therefore, it is best to remove it all together. This will simplify the code a little bit and also bring performance improvements as there will be fewer thread synchronization points during the graph computation

Related:

• Numa llama.cpp#1556

[rlanday]

Hi, I took a look at this and found that GGML_TASK_FINALIZE is used by ggml_compute_forward_cross_entropy_loss_f32():

https://github.com/ggerganov/llama.cpp/blob/178b1850ebd21b349cebbee887950e435c5aa2d3/ggml.c#L15453

Please advise.