sycl: Fix and disable more configurations of mul_mat

[Rbiessy]

Follow up of PR #15092

The previous PR missed to fix some configurations.
This PR fixes one case with oneDNN but disabled another one (type_a=f16,type_b=f32,m=1056,n=1,k=129,bs=[8,3],nr=[4,1],per=[0,2,1,3],v=0). It should be possible to support it with the current oneDNN logic here but I'm not able to find a solution in a timely manner. I don't expect I will be able to work on it myself so feel free to continue this fix.

This should be enough to make the SYCL CI green again. The PR also fixes some compilation warnings.

[qnixsynapse]

Still failing this case:

[MUL_MAT] NMSE = 0.636762990 > 0.000500000   MUL_MAT(type_a=f16,type_b=f32,m=1057,n=1,k=129,bs=[1,3],nr=[4,1],per=[0,2,1,3],v=0): FAIL

[Rbiessy]

That's weird, I can definitely see this test passing on LunarLake iGPU and on PVC. Would you be able to have a closer look on your side?

[qnixsynapse]

Yeah double checked. Failing exactly here:

[MUL_MAT] NMSE = 0.667358089 > 0.000500000   MUL_MAT(type_a=f16,type_b=f32,m=1057,n=1,k=129,bs=[1,3],nr=[4,1],per=[0,2,1,3],v=0): FAIL

Unfortunately, I do not have enough time to debug this. And thank you for taking a look into it.

[Rbiessy]

Are you using oneDNN? The other path may still be broken. The next step should be to remove oneMKL and oneMath so we may want to start disabling more configurations when oneDNN is not used if that's the reason why it's failing for you.

[qnixsynapse]

Yes, it's the default using oneDNN