[LinalgExt] Added toggle for using useExp2 for onlineAttention Decomposition

[keshavvinayak01]

Following the discussion from #22441

Depending on the backend, certain computations may benefit from directly using exp instead of exp2, since there might be accuracy losses due to FP-reassociation. It's helpful to add flag incase the user tracks losses to this particular computation and might favour directly using exp.

The use_exp2 flag is mostly unused in dialect conversions and passes, I presume it's used as a KernelOption. The changes here will not modify the default behavior.

[MaheshRavishankar]

I think it would be simpler to just add an optional attribute to the op itself to use exp2 for decomposition. This doesnt need to be a separate attribute that is not part of the op definition.

[keshavvinayak01]

I think it would be simpler to just add an optional attribute to the op itself to use exp2 for decomposition. This doesnt need to be a separate attribute that is not part of the op definition.

I was going to do that, but then I saw

iree/compiler/src/iree/compiler/Codegen/Common/test/strip_compilation_info.mlir

Line 101 in 70b2b45

%result = iree_linalg_ext.attention {decomposition_config = {pv_attrs = {x}, qk_attrs = {y}, use_exp2}, indexing_maps = [#map, #map1, #map2, #map3, #map4], compilation_info = #compilation} ins(%arg0, %arg1, %arg2, %arg3 : tensor<2x10x6x4xf16>, tensor<2x10x4x4xf16>, tensor<2x10x4x4xf16>, f16) outs(%init : tensor<2x10x6x4xf16>) {

Where it's part of the decomposition config itself. So I thought I'd better refine the attribute and continue using it. Making it an optional attribute of the op itself might introduce redundancy since this already exists @MaheshRavishankar ?