Add type conversion to convert data types to supported data types in ttnn

[mtopalovicTT]

ttnn doesn't support all data types which come from forge. One example is is int32 is mapped to Uint32
This can lead to some strange IR like this

to_layout(%arg1) { dtype = Uint32 }(tensor<10x10xi32) -> tensor<10x10xi32)

what happens is that during lowering to ttnn when creating dtype attribute for to_layout we convert outputTensor data type to ttnn supported data type. More over IR doesn't correctly represents what is happening in runtime since we would either way convert all types from IR into ttnn supported types during translation to flatbuffer....

One approach to do this is to do type conversion on input graph from forge. This would update all types to ttnn supported types.
Maybe this is okay from forge perspective since they don't do anything on MLIR graph so if we change type in IR it wouldn't make difference on them.
Different approach is to change forge to do type conversion to dtypes which are supported in ttnn. This would make more sense since both input IR and output IR from mlir would have identical data types.

[mtopalovicTT]

cc @svuckovicTT @nsmithtt @pilkicTT

[nsmithtt]

@mtopalovicTT, I think TTIR should be flexible and accept non-hw types like ints and bools and complex types etc. We should handle this kind of data format conversion during a TTIR->TTIR pass.

@mrakitaTT, @AleksKnezevic, any input from the stablehlo side? There have been many similar df conversion situations there.

[AleksKnezevic]

Yeah, there were a few cases where we've had to convert unsupported types. @mmanzoorTT did the conversion, so he can comment more.