CUDA: fix MMQ stream-k fixup ne1 indices

[JohannesGaessler]

See discussion starting with ikawrakow/ik_llama.cpp#728 (comment) , the use of MMQ MoE optimizations is resulting in increased perplexity on master. The problem is that the wrong indices are being used when determining which dst columns should be receiving the stream-k fixup. This is a very typical bug that I encounter during development but unfortunately? this is one of the rare cases where the impact is small enough to be overlooked. Generally speaking, the impact will be largest for the combination of small models are large GPUs where the SM count is not a power of 2 (RTX 4090 for example has 128 SMs). Example models:

Model	GPU	PPL master	PPL PR
GraniteMoe 3b q4_0	RTX 3090	10.5577	10.0799
GraniteMoe 3b q4_0	RTX 4090	10.0879	10.0877
GraniteMoe 3b q4_0	RTX 5090	15.4910	10.0853
Qwen 3 30b q4_0	RTX 3090	9.3418	9.2930
Qwen 3 30b q4_0	RTX 4090	9.2928	9.2928
Qwen 3 30b q4_0	RTX 5090	9.4052	9.2930

The bug in question has been on master since #13199 though the impact became larger with #15525 when the upper bound for CUDA blocks was tightened and more stream-k seams ended up in tiles that are not being skipped.