ttnn.where (and eltwise ops in general) mixed data types issue

[azecevicTT]

As title says issue is related to behaviour of (binary) eltwise ops in the presence of mixed data type operands. Correct me if I'm wrong but currently the result's data type will default to the data type of the first operand, which leads to some weird cases like:

ttnn.add(bf16, f32) -> bf16
ttnn.add(f32, bf16) -> f32

Most frameworks would default to f32 in both cases. Is there an initiative to make eltwise ops work like this in TTNN?

Related to that is the behaviour of ttnn.where. The data type of the result will always be the data type of condition operand, which is a bit weird, and not what most people would expect. Just by switching the order of condition and value https://github.com/tenstorrent/tt-metal/blob/main/ttnn/cpp/ttnn/operations/eltwise/ternary/where.cpp#L37 leads to more expected result (though it might cause some other issues). I believe ttnn.where is a bit special in that regard, even if you want to make case for strongly typed operations where mixed types are considered errors, while it's reasonable to assume true_value and false_value have the same data type, I believe that condition's data type should be independent of them (and consequently the result's data type).

[AleksKnezevic]

@cmaryanTT, could we get an ETA?