02-fused-softmax: PyTorch faster than Triton on Max 1550 in composite mode

[pbchekin]

ZE_FLAT_DEVICE_HIERARCHY=FLAT

ZE_FLAT_DEVICE_HIERARCHY=COMPOSITE

[chengjunlu]

I think it maybe caused by the register spill.
In the composite mode, the private space for spilling maybe always allocated in one tile and all the threads on the other tile has to access the private space across the EMIB bus.

We may need to implement a new softmax kernel which is close to the torch's implementation and re-test the performance.

[anmyachev]

@pbchekin in composite mode we need to adjust cache size (double it) #2265 (comment). I wonder how the chart will change after this change.