CUDA: fix unpadded strides in MMA FA kernel

[JohannesGaessler]

Fixes #17796 .

The problem is that for determining whether the kernel can assume aligned and padded allocations the code on master is only considering the tensor shapes, but not the tensor strides. So with the bandwidth-optimal copy sizes of 4/16 bytes you can end up with misaligned pointers.

Performance changes

GPU	Model	Microbatch size	Test	t/s 6b82eb7	t/s d4cbe79	Speedup
RTX 3090	llama 7B Q4_0	1	pp512@d32768	40.05	40.02	1.00
RTX 3090	llama 7B Q4_0	2	pp512@d32768	76.19	75.11	0.99
RTX 3090	llama 7B Q4_0	4	pp512@d32768	146.70	144.65	0.99
RTX 3090	llama 7B Q4_0	8	pp512@d32768	250.73	247.66	0.99
RTX 3090	llama 7B Q4_0	16	pp512@d32768	504.32	503.10	1.00
RTX 3090	llama 7B Q4_0	32	pp512@d32768	905.59	907.51	1.00
RTX 3090	llama 7B Q4_0	64	pp512@d32768	1475.74	1473.77	1.00
RTX 3090	llama 7B Q4_0	128	pp512@d32768	1567.34	1573.62	1.00
RTX 3090	llama 7B Q4_0	256	pp512@d32768	1598.65	1557.23	0.97
RTX 3090	llama 7B Q4_0	512	pp512@d32768	1432.27	1389.90	0.97
RTX 3090	llama 8B Q4_0	1	pp512@d32768	93.02	92.98	1.00
RTX 3090	llama 8B Q4_0	2	pp512@d32768	170.50	170.45	1.00
RTX 3090	llama 8B Q4_0	4	pp512@d32768	304.23	304.14	1.00
RTX 3090	llama 8B Q4_0	8	pp512@d32768	439.36	440.12	1.00
RTX 3090	llama 8B Q4_0	16	pp512@d32768	874.90	874.54	1.00
RTX 3090	llama 8B Q4_0	32	pp512@d32768	1199.90	1197.56	1.00
RTX 3090	llama 8B Q4_0	64	pp512@d32768	1507.73	1504.91	1.00
RTX 3090	llama 8B Q4_0	128	pp512@d32768	1646.17	1641.61	1.00
RTX 3090	llama 8B Q4_0	256	pp512@d32768	1775.90	1770.39	1.00
RTX 3090	llama 8B Q4_0	512	pp512@d32768	1545.68	1550.91	1.00
RTX 4090	llama 7B Q4_0	1	pp512@d32768	43.75	43.75	1.00
RTX 4090	llama 7B Q4_0	2	pp512@d32768	84.11	83.30	0.99
RTX 4090	llama 7B Q4_0	4	pp512@d32768	167.64	166.11	0.99
RTX 4090	llama 7B Q4_0	8	pp512@d32768	314.29	312.07	0.99
RTX 4090	llama 7B Q4_0	16	pp512@d32768	606.69	605.35	1.00
RTX 4090	llama 7B Q4_0	32	pp512@d32768	1140.19	1140.27	1.00
RTX 4090	llama 7B Q4_0	64	pp512@d32768	2043.45	2043.16	1.00
RTX 4090	llama 7B Q4_0	128	pp512@d32768	3146.05	3127.82	0.99
RTX 4090	llama 7B Q4_0	256	pp512@d32768	3152.41	3136.11	0.99
RTX 4090	llama 7B Q4_0	512	pp512@d32768	2563.87	2555.08	1.00
RTX 4090	llama 8B Q4_0	1	pp512@d32768	104.92	104.94	1.00
RTX 4090	llama 8B Q4_0	2	pp512@d32768	190.14	190.76	1.00
RTX 4090	llama 8B Q4_0	4	pp512@d32768	367.05	367.31	1.00
RTX 4090	llama 8B Q4_0	8	pp512@d32768	647.03	646.37	1.00
RTX 4090	llama 8B Q4_0	16	pp512@d32768	1142.95	1142.36	1.00
RTX 4090	llama 8B Q4_0	32	pp512@d32768	2067.27	2068.11	1.00
RTX 4090	llama 8B Q4_0	64	pp512@d32768	3155.18	3153.10	1.00
RTX 4090	llama 8B Q4_0	128	pp512@d32768	3516.50	3512.25	1.00
RTX 4090	llama 8B Q4_0	256	pp512@d32768	3260.76	3247.34	1.00
RTX 4090	llama 8B Q4_0	512	pp512@d32768	2679.73	2662.31	0.99
RTX 5090	llama 7B Q4_0	1	pp512@d32768	74.51	74.51	1.00
RTX 5090	llama 7B Q4_0	2	pp512@d32768	131.12	129.75	0.99
RTX 5090	llama 7B Q4_0	4	pp512@d32768	257.78	255.16	0.99
RTX 5090	llama 7B Q4_0	8	pp512@d32768	467.28	463.29	0.99
RTX 5090	llama 7B Q4_0	16	pp512@d32768	837.73	881.04	1.05
RTX 5090	llama 7B Q4_0	32	pp512@d32768	1627.55	1602.04	0.98
RTX 5090	llama 7B Q4_0	64	pp512@d32768	2623.73	2551.33	0.97
RTX 5090	llama 7B Q4_0	128	pp512@d32768	2803.75	2771.09	0.99
RTX 5090	llama 7B Q4_0	512	pp512@d32768	2684.81	2525.63	0.94
RTX 5090	llama 8B Q4_0	1	pp512@d32768	165.94	165.86	1.00
RTX 5090	llama 8B Q4_0	2	pp512@d32768	269.26	269.14	1.00
RTX 5090	llama 8B Q4_0	4	pp512@d32768	486.58	486.46	1.00
RTX 5090	llama 8B Q4_0	8	pp512@d32768	810.03	809.42	1.00
RTX 5090	llama 8B Q4_0	16	pp512@d32768	1405.56	1404.92	1.00
RTX 5090	llama 8B Q4_0	32	pp512@d32768	2330.73	2331.65	1.00
RTX 5090	llama 8B Q4_0	64	pp512@d32768	2866.00	2861.02	1.00
RTX 5090	llama 8B Q4_0	128	pp512@d32768	3502.23	3495.91	1.00
RTX 5090	llama 8B Q4_0	256	pp512@d32768	3370.96	3370.50	1.00
RTX 5090	llama 8B Q4_0	512	pp512@d32768	2922.97	2929.01	1.00
V100-PCIE-32GB	llama 7B Q4_0	1	pp512@d32768	37.36	37.36	1.00
V100-PCIE-32GB	llama 7B Q4_0	2	pp512@d32768	68.72	68.70	1.00
V100-PCIE-32GB	llama 7B Q4_0	4	pp512@d32768	126.26	126.18	1.00
V100-PCIE-32GB	llama 7B Q4_0	8	pp512@d32768	222.84	223.12	1.00
V100-PCIE-32GB	llama 7B Q4_0	16	pp512@d32768	332.87	333.29	1.00
V100-PCIE-32GB	llama 7B Q4_0	32	pp512@d32768	524.77	594.61	1.13
V100-PCIE-32GB	llama 7B Q4_0	64	pp512@d32768	437.89	460.85	1.05
V100-PCIE-32GB	llama 7B Q4_0	128	pp512@d32768	634.31	679.05	1.07
V100-PCIE-32GB	llama 7B Q4_0	256	pp512@d32768	857.35	842.09	0.98
V100-PCIE-32GB	llama 7B Q4_0	512	pp512@d32768	949.65	921.29	0.97
V100-PCIE-32GB	llama 8B Q4_0	1	pp512@d32768	81.32	81.50	1.00
V100-PCIE-32GB	llama 8B Q4_0	2	pp512@d32768	153.67	154.03	1.00
V100-PCIE-32GB	llama 8B Q4_0	4	pp512@d32768	220.77	220.32	1.00
V100-PCIE-32GB	llama 8B Q4_0	8	pp512@d32768	311.82	311.82	1.00
V100-PCIE-32GB	llama 8B Q4_0	16	pp512@d32768	471.04	470.87	1.00
V100-PCIE-32GB	llama 8B Q4_0	32	pp512@d32768	671.06	671.55	1.00
V100-PCIE-32GB	llama 8B Q4_0	64	pp512@d32768	465.56	465.20	1.00
V100-PCIE-32GB	llama 8B Q4_0	128	pp512@d32768	697.67	696.30	1.00
V100-PCIE-32GB	llama 8B Q4_0	256	pp512@d32768	857.68	856.99	1.00
V100-PCIE-32GB	llama 8B Q4_0	512	pp512@d32768	935.99	932.80	1.00

For most models the performance should remain unchanged since they use GQA and this PR only affects models without GQA or if insufficient padding is used. I'm not sure what is happening with the V100 performance at medium batch sizes - I didn't yet really optimize this kernel for that particular GPU so I'm wondering whether the new loop structure results in different register usage.

[CISC]

https://github.com/ggml-org/llama.cpp/actions/runs/20079357666/job/57602187703?pr=17891#step:3:23872