Offload and modules with unused submodules

[sgugger]

This PR deals with models with unused submodules combined with offload (on CPU and disk). By unused submodules, we mean a something like the ModuleWithUnusedSubModules introduced in the test file in this PR: a module that defines a submodule, but does not call it during its forward, just does some operations using its weights.

Currently, the offload will fail for such modules because the weights are fetched in a pre-forward hook and the forward method is never called. This PR fixes that by introducing a new argument named load_all_weights_classes (if you have any better suggestion, please let me know as I'm not super happy with this name) which contains the name of such submodules used (like no_split_module_classes contains the names of the modules that shouldn't be split across devices).

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[muellerzr]

Thanks! The fix makes sense to me, left one naming suggestion/alternative

[pacman100]

In attach_align_device_hook_on_block, I think below change is needed

if not isinstance(execution_device, Mapping):
-        execution_device = {key: offload for key in offload.keys()}
+        execution_device = {key: execution_device for key in offload.keys()}

Left a minor suggestion too.
Apart from that everything LGTM!