POC: combined scale + diagonal mask infinity + soft max op

[ikawrakow]

TL;DR: This PR is a POC that combines scale + diagonal mask infinity + soft max, used on K * Q in the attention part of the network, to raise the question if we want to get involved with fusing operations.

It is a POC, so for now just a simple Metal implementation (i.e., none of the optimizations to these operations in #3084 are used). Despite this, fusing these 3 operations leads to a small, but measurable performance gain (see table). If the consensus is that this is a worthwhile thing to do, I can add implementation for the other backends, and would see if it is possible to optimize the kernel(s) for the combined scale + diagonal mask infinity + soft max operation.

model	backend	test	t/s (Master)	t/s (PR)	Speedup
Falcon 7B mostly F16	Metal	pp 512	379.71 ± 0.23	386.22 ± 0.43	1.017
LLaMA 7B mostly F16	Metal	pp 512	539.25 ± 0.30	543.57 ± 0.35	1.008
LLaMA 7B mostly Q4_0	Metal	pp 512	495.43 ± 0.59	499.20 ± 0.35	1.008
Falcon 7B mostly F16	Metal	tg 128	23.26 ± 0.01	23.48 ± 0.05	1.009
LLaMA 7B mostly F16	Metal	tg 128	24.14 ± 0.04	24.34 ± 0.10	1.008
LLaMA 7B mostly Q4_0	Metal	tg 128	62.82 ± 0.01	63.91 ± 0.04	1.017

[ikawrakow]

Given the complete lack of interest, I'm closing this one.