Add sharding tests to multigpu-test.sh and fix custom operator decorator (#77825)
1. Enabled multigpu tests.
2. Fixed failing multigpu tests.
3. Fixed custom operator decorator to be first preference in operator dispatch.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/77825
Approved by: https://github.com/wanchaol, https://github.com/fduwjj
