fix: memory cycle in diverts

[brentjanderson]

There is an Rc<T> cycle between a Container, a Divert, back to a Container. This reliably produces a memory leak when running multiple stories in sequence or parallel (memory_leak_test helps verify this).

By caching potential targets and using Weak<T> to break the cycle, we avoid this memory leak case.

This shouldn't change the performance profile of a running story, as target references are already lazily loaded. We are just moving from the Divert struct to a shared cache with weak references.

[bladecoder]

Thank you, @brentjanderson, for identifying this real reference-cycle leak and for proposing the correct core solution: caching the target container through Weak instead of retaining it through Rc. Your diagnosis and approach have been preserved, and you are credited as a co-author of the replacement change.

I am closing this PR because it can no longer be merged cleanly in its current form:

• the repository has since moved the runtime crate from lib/ to runtime/, leaving this branch with merge conflicts;
• the nested if let in the patch now fails the repository’s mandatory Clippy run with -D warnings (clippy::collapsible-if);
• the regression test installs a process-wide unsafe tracking allocator and runs 1,000 complete stories. Locally it took roughly 9–13 seconds and, when demonstrating the regression, retained about 667 MB. That makes a failing test unnecessarily expensive and risks an OOM on constrained CI runners.

The replacement PR, #32, ports the fix to the current tree, satisfies the current lint rules, and replaces the allocator-based stress test with a focused ownership test that constructs the exact Container -> Divert -> Container cycle and verifies deallocation directly through Weak. It has passed CI and has now been merged.

Thanks again for the valuable report and original fix.