Improve type constraints for aievec:mul_elem and Co.

[david-vc]

Conversion passes like AIEVecToLLVM rely on invariants concerning the number of lanes of operands and results, but these invariants are currently spread out over the code base for operations like aievec.mul_elem. It would be desirable to consolidate the invariants like it has been done for for example aievec.matmul.