Support sharding in PyTorchHubMixin

[Wauplin]

Depends on #2065.

[not-lain]

@Wauplin
I think that #1995 by @NielsRogge has some interesting resources and maybe we can use that as a reference for this issue.

[Wauplin]

Actually #1995 was a start but I'd prefer to address #2065 first to make the logic reusable. I can take care of it.

Just to be sure in term of priorities, have we already encountered a library where we'd like to shard the weights or not yet?

[not-lain]

@Wauplin not yet, all the AI models that i've worked on so far are small (<1GB) so all good so far

[Wauplin]

Thanks for confirming! Doesn't mean we shouldn't handle it but at least it's not critical :)