test: elementwise ops on mismatched sizes throw DimensionMismatch

[JohnCobbler]

Adding mismatched-size GPU arrays with + used to silently truncate to the
shorter operand (JuliaGPU/CUDA.jl#2812). That path is gone — + now goes
through Base promote_shape and throws DimensionMismatch symmetrically,
matching Base.Array. This adds a guard in the broadcasting testsuite so it
stays that way.

The check sits in the parameterized testsuite, so it runs for every backend;
on CPU CI it exercises the same dispatch (the throw happens in promote_shape
before any kernel launch), and a CUDA backend confirms the device path.