GitLab

When testing gpu::contiguous instruction replacability recursive calls to try_compute_shape can pick up gpu::contiguous instructions as well.
These instructions always passes the try_compute_shape check because compute_shape of gpu::contiguous always returns a standard shape, thus propagating false `true` values upwards.
For these corner cases we must demand the standard shape of the inputs as well, to avoid ellimination of needed gpu::contiguous instructions.