metal : use residency sets

[ggerganov]

fix #10119

Using residency sets makes the allocated memory stay wired and eliminates almost completely the overhead observed in #10119. For example, on M2 Ultra, using 7B Q8_0 model the requests are ~250ms faster thanks to this change. It seems it is not necessary to attach the residency sets to the command queue and buffers, so the change is rather simple. For each buffer, we create an associated MTLResidencySet and add the MTLBuffer objects to it. After that we commit it and request residency:

/~https://github.com/ggerganov/llama.cpp/blob/225d2e0ca1d7a7e627f2cea4a43dd77a83b9f078/ggml/src/ggml-metal/ggml-metal.m#L1084-L1091

build: b9126fe (4561)

Model	Test	t/s master	t/s gg/metal-residency-sets	Speedup
llama 3B F16	pp512	3289.51	3286.29	1.00
llama 3B F16	tg128	73.28	73.35	1.00
llama 3B Q4_0	pp512	2999.71	3002.93	1.00
llama 3B Q4_0	tg128	165.83	166.03	1.00
llama 3B Q8_0	pp512	2958.32	2960.69	1.00
llama 3B Q8_0	tg128	123.61	123.96	1.00

Metal backend changes

Checks the environment variable GGML_METAL_NO_RESIDENCY. If set, then no residency sets will be created, allowing the GPU memory to be collected by the OS after 1 second of inactivity. Generally, this should rarely be needed as it hurts the performance of the application, but keeping support just in case.

[ggerganov]

Great news - this change finally resolves the annoying overhead that I was observing. The only remaining question is how to implement this to be compatible with macOS < 15.0.

Any suggestions?

Edit: resolved