CUDA: fix MMQ stream-k for --split-mode row #8167

[Nexesenex]

Courtesy of Johannes Gaessler, as usual. PR came after refactor, but I think that it's important to have it merged now for those using split row.

https://github.com/ggerganov/llama.cpp/pull/8167

• [x] I have read the contributing guidelines
• Self-reported review complexity:
  • [x] Low
  • [ ] Medium
  • [ ] High

[LostRuins]

Will merge the commit from upstream directly