DiffusionGemma: mask layout and CI (#46654)
* try
* cache draft
* some todos in modular
* oops
* .
* fix
* fix repo
* this is not fun
* should we force cuda, the runners seem to be falling back on cpu ?
* this
* address some commnets
* oops
* what if
* revert as well
* no cache append
* revert
* cant trust claude
