Cross-thread backward: forward on one thread, .backward() on another fails with a misnamed 'second time' error #78

Found by the free-threading audit for #51 (item R3), measured on the 3.13 shim with the GIL on. It is not a free-threading-only issue.

x = torch.ones(3, requires_grad=True)
# forward in a worker thread
y = (x * 3).sum()          # computed in threading.Thread
y.backward()               # main thread

• Upstream torch 2.13: x.grad == tensor([3., 3., 3.]).
• Shim: RuntimeError: torch._C eager: Trying to backward through the graph a second time -- the saved intermediate values have already been freed ...

Cause. The eager tape (EAGER in torch_c/src/capture.rs:66-75) is a thread_local. The backward thread finds no tape and reports graph reuse. Upstream's graph lives on the tensors, so any thread can walk it.

This is wrong twice over:

• the shim refuses something upstream does;
• the refusal names the wrong cause (AGENTS.md §18).

Done when

• A cross-thread forward/backward gives upstream's gradient. The likely fix is that the tape becomes reachable from the tensors' grad_fn rather than from the calling thread, or is process-wide under a lock.
• Until that lands: if the tape is missing because the forward ran on another thread, the error says that, not "a second time".
• A test runs the forward in a thread and the backward in another, compared against upstream.