Fix forwarding/move bug (#53556)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/53556
When packing a `Tensor&` (mutable lvalue reference) into an IValue, we accidentally didn't increase the refcount.
This wasn't triggered anywhere, until I tried to enable backend fallbacks. Backend fallbacks for ops that
have out arguments (i.e. ops that take `Tensor&` arguments and return `Tensor&` arguments) pack those returns
into an IValue stack (and accidentally don't increase the refcount), then later that stack gets destructed
(which decreases the refcount and possibly destroys the Tensor), and the `Tensor&` passed in as an out argument
is suddenty freed memory.
This PR fixes that by forwarding instead of moving when wrapping Tensors into IValues.
ghstack-source-id: 125886986
(Note: this ignores all push blocking failures!)
Test Plan: waitforsandcastle
Reviewed By: swolchok
Differential Revision: D26896507
fbshipit-source-id: 62102fa89e522699b5174c33279a2b1a775066a4
