[ROCM] Gibberish on wave64 devices

[IMbackK]

It would seem like exllamav2 do sent work on device with a wave size of 64, the output is gibberish, i tested this on a MI25, MI50 and MI100 all of which are wave64 devices.
At the same time wave32 devices such as rx6800xt work fine.
exllama 1 works fine in all cases.

I tested this on ROCM 5.7 with pythoch 2.1

[IMbackK]

dupe of #33 , indeed it is warp layout