fix `hash_storage`'s padding calculation (#105036)
Fixes #105035.
The existing implementation attempts to make `x.numel() % 4 == 0` by appending `x.numel() % 4` zeros. This is backwards, e.g if `x.numel() % 4 == 1`, we need to append `[0, 0, 0]`, not `[0]`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/105036
Approved by: https://github.com/soulitzer, https://github.com/ezyang