CUDA: fix kernel selection logic for tile FA

[JohannesGaessler]

Fixes #19652 .

The problem is that the kernel selection logic is slightly wrong so the host code runs into an abort.

[pwilkin]

@JohannesGaessler CI seems to dislike something:

  MUL_MAT(type_a=q5_1,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1)
  Backend CUDA0: FAIL

[JohannesGaessler]

2026-02-17T14:06:08.1940247Z [MUL_MAT] ERR = 0.000589984 > 0.000500000   MUL_MAT(type_a=q5_1,type_b=f32,m=16,n=1,k=256,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1): �[1;31mFAIL�[0m

The failing test is a matrix multiplication which is 100% unaffected by the changes in this PR. Since we are using random seeds in the tests it can happen from time to time that the small differences between the CUDA and CPU backends randomly exceed the tolerance we set, particularly if the input matrices are small and the differences are less likely to average out.

[ggerganov]

Adding a comment/reference of the DV limit would be useful.

[pwilkin]

The failing test is a matrix multiplication which is 100% unaffected by the changes in this PR. Since we are using random seeds in the tests it can happen from time to time that the small differences between the CUDA and CPU backends randomly exceed the tolerance we set, particularly if the input matrices are small and the differences are less likely to average out.

Aight, just throwing out an idea: maybe it would be a good idea to vary the permitted NMSE divergence according to tensor size, to avoid false positives like this?