Add option to disable MMA support on Turing

[pt13762104]

This helps with the performance of GTX 16 series (Turing without tensor cores), since it seems like code that uses tensor cores don't execute at nearly the full speed.

Perplexity

Original	This PR
11.6842 +/- 0.09676	11.6837 +/- 0.09676

Performance (with GGML_CUDA_FORCE_MMQ)

Original (no GGML_CUDA_FORCE_MMQ)

model	size	params	backend	ngl	fa	test	t/s
qwen3 4B Q8_0	3.98 GiB	4.02 B	CUDA	99	0	pp512	249.20 ± 0.05
qwen3 4B Q8_0	3.98 GiB	4.02 B	CUDA	99	1	pp512	260.86 ± 3.81

Original

model	size	params	backend	ngl	fa	test	t/s
qwen3 4B Q8_0	3.98 GiB	4.02 B	CUDA	99	0	pp512	261.56 ± 1.26
qwen3 4B Q8_0	3.98 GiB	4.02 B	CUDA	99	1	pp512	282.52 ± 0.04

Original (CMAKE_CUDA_ARCHITECTURES=61)

model	size	params	backend	ngl	fa	test	t/s
qwen3 4B Q8_0	3.98 GiB	4.02 B	CUDA	99	0	pp512	633.76 ± 1.58
qwen3 4B Q8_0	3.98 GiB	4.02 B	CUDA	99	1	pp512	840.55 ± 4.35

This PR (no GGML_CUDA_FORCE_MMQ)

model	size	params	backend	ngl	fa	test	t/s
qwen3 4B Q8_0	3.98 GiB	4.02 B	CUDA	99	0	pp512	796.90 ± 1.07
qwen3 4B Q8_0	3.98 GiB	4.02 B	CUDA	99	1	pp512	844.69 ± 3.31

This PR

model	size	params	backend	ngl	fa	test	t/s
qwen3 4B Q8_0	3.98 GiB	4.02 B	CUDA	99	0	pp512	918.69 ± 2.68
qwen3 4B Q8_0	3.98 GiB	4.02 B	CUDA	99	1	pp512	985.77 ± 2.17

[JohannesGaessler]

You should be able to get the same effect on master by compiling with CMAKE_CUDA_ARCHITECTURES=61. I think it would make more sense to check the names of available GPUs and print a warning if GTX 1600 is used with code compiled for CC 7.0 or 7.5. It will not be possible to use GTX 1600 optimally in conjunction with other Turing GPUs but that would require a more general refactor of the code and would I think be a lot of complexity for a setup that basically no one will ever use.

[pt13762104]

With -DCMAKE_CUDA_ARCHITECTURES=61 -DGGML_CUDA_FORCE_MMQ=1, I get the performance roughly equal to 36201c6.
If I disable GGML_CUDA_FORCE_MMQ, performance instantly drops to 0.7 TFLOPS (see below for Llama 2 results):

model	size	params	backend	ngl	fa	test	t/s
llama 7B Q4_0	3.56 GiB	6.74 B	CUDA	99	0	pp512	47.15 ± 0.18
llama 7B Q4_0	3.56 GiB	6.74 B	CUDA	99	0	tg128	44.60 ± 0.27
llama 7B Q4_0	3.56 GiB	6.74 B	CUDA	99	1	pp512	47.10 ± 0.72
llama 7B Q4_0	3.56 GiB	6.74 B	CUDA	99	1	tg128	43.74 ± 0.75