Add NVIDIA Turing support to FlashMMA kernel

[nlzy]

Summary

The flash attention kernel in cpp/neuralnet/cudaflashmma.cuh needed sm_80+, because it used mma.sync.aligned.m16n8k16 and cp.async.

On NVIDIA Turing GPUs (sm_75), transformer models fallback to a much slower attention path. cuDNN graph SDPA also needs sm_80, so it is not available either.

This PR added Turing support to cudaflashmma.cuh, with a +70.3% speedup compared to original CUDA backend on transformer models.

Environment

• Intel Xeon CPU E5-2620 v4 @ 2.10GHz
• NVIDIA RTX 2080 Ti (22GB mod)
• Ubuntu 24.04 with Linux 6.8 kernel
• CUDA 12.9, cuDNN 9.10.2, TensorRT 10.12.0
• Model: kata1-tf3-b11c768-s11001M-d5973M.bin.gz

Testing

• kata-raw-nn on the empty board (komi 7.5), mma on vs off: whiteWin 0.655850 vs 0.656328, whiteLead 0.876 vs 0.882.
• GTP genmove sanity check passes.
• ./katago runtests: all pass.
• ./katago runnnlayertests: all pass.
• cpp/rungpuerrortest.sh: all pass.
• cpp/runcudaopttests.sh with EXPECT_FUSED_FFN=0 EXPECT_SDPA=0: all pass.

Performance

./katago benchmark -config gtp.cfg -model <tf3 model> -v 5000 -n 5

CUDA builds use numNNServerThreadsPerModel = 2. TensorRT use 1.

Build	Best nnEvals/s	visits/s	Config (batch / threads)	vs Origin (CUDA)
Origin (CUDA)	400.13	661.56	16 / 32	—
Origin (TRT)	589.89	994.32	16 / 32	+47.4%
This PR (CUDA)	681.60	1161.40	16 / 32	+70.3%

This PR is also +15.5% faster than Origin TensorRT backend in nnEvals/s.

Full sweep (nnEvals/s):

Build	batch	t=8	t=16	t=24	t=32
Origin (CUDA)	8	327.11	383.63	391.62	391.25
Origin (CUDA)	16	325.12	376.99	394.48	400.13
Origin (CUDA)	32	324.59	375.04	388.73	395.49
Origin (TRT)	8	466.90	542.44	541.67	542.60
Origin (TRT)	16	488.71	544.74	569.59	589.89
Origin (TRT)	32	484.76	551.48	569.79	581.18
This PR (CUDA)	8	564.38	641.43	657.90	654.25
This PR (CUDA)	16	556.40	624.94	659.96	681.60
This PR (CUDA)	32	556.75	623.48	652.76	672.68

[lightvector]

Thanks!