RFC: remove arith/math ops on tensors

[kuhar]

@stellaraccident

• Static to dynamic: there is an argument here that this should be allowed, but I would side on this trait being strict and disallowing.
• Element type change, implying a size/indexing change. This is actually some form of bitcast where the result should be tensor<2xi64> if it were allowed. This was never intended to be allowed by the ElementwiseMappable trait and is illegal IR, afaict.

To clarify, the point I was trying to make is that the first cast is surprising to me, not casting over tensors in general. IE:

• arith.extsi %x, %y : tensor<12xi16> to tensor<12xi64> → fine
• arith.extsi %x, %y : tensor<?xi16> to tensor<?xi64> → fine
• arith.extsi %x, %y : tensor<12xi16> to tensor<?xi64> → not fine

@ThomasRaoux

+1 to what has been said I would also expect it to be elementwise meaning the input and output shape should match in IR.

Yep. So I think everyone is on the same page here so far.