[functorch] Add _unsafe_set_level for torchdim (pytorch/functorch#924)
[torchdim](https://github.com/facebookresearch/torchdim) reuses BatchTensorImpl objects
where it can across batched invocations, but needs to patch their levels to match
what _vmap_increment_nesting returns.
[ghstack-poisoned]
