[mlir][tosa] Enable decomposing Conv2D also where 1 input dim is dynamic

Restricted to just 1 dynamic input dim as that worked all the way through to
codegen.

Differential Revision: https://reviews.llvm.org/D129334