OpenCL: add initial FA support

[rmatif]

This PR introduces F16/F32 FA support for the OpenCL backend. It has been extremely challenging to achieve good performance on this kind of hardware, but I believe it is now decent enough to serve as a baseline that we can further iterate on. I also believe there is room for improvement for tg

Results on Adreno 830:

model	size	params	backend	ngl	fa	test	t/s
llama 1B F16	2.30 GiB	1.24 B	OpenCL	99	0	pp512	198.69 ± 0.59
llama 1B F16	2.30 GiB	1.24 B	OpenCL	99	0	tg128	21.88 ± 0.85
llama 1B F16	2.30 GiB	1.24 B	OpenCL	99	1	pp512	274.75 ± 1.22
llama 1B F16	2.30 GiB	1.24 B	OpenCL	99	1	tg128	21.58 ± 0.39

Adreno 750:

model	size	params	backend	ngl	fa	test	t/s
llama 1B F16	2.30 GiB	1.24 B	OpenCL	99	0	pp512	139.96 ± 0.51
llama 1B F16	2.30 GiB	1.24 B	OpenCL	99	0	tg128	19.70 ± 0.11
llama 1B F16	2.30 GiB	1.24 B	OpenCL	99	1	pp512	151.22 ± 0.85
llama 1B F16	2.30 GiB	1.24 B	OpenCL	99	1	tg128	17.94 ± 0.15

[lhez]

@rmatif Very cool, thank you!

[lhez]

Sorry, got distracted during the past week. Will come back to this asap.