Replace @onlyCUDA with @onlyAccelerator on 4 tests so they run on
all accelerator backends (CUDA, MPS, XPU, etc.):

test_set_default_dtype_works_with_foreach: replace @onlyCUDA with
@onlyAccelerator, add @skipMPS (MPS does not support float64)
test_fused_does_not_step_if_foundinf: replace @onlyCUDA with
@onlyAccelerator, fix device string matching...